Benchmarking weakly-supervised deep learning pipelines for whole slide classification in computational pathology

Automated curation of large-scale cancer histopathology image datasets using deep learning

BACKGROUND

Artificial intelligence (AI) has numerous applications in pathology, supporting diagnosis and prognostication in cancer. However, most AI models are trained on highly selected data, typically one tissue slide per patient. In reality, especially for large surgical resection specimens, dozens of slides can be available for each patient. Manually sorting and labelling whole-slide images (WSIs) is a very time-consuming process, hindering the direct application of AI on the collected tissue samples from large cohorts. In this study we addressed this issue by developing a deep-learning (DL)-based method for automatic curation of large pathology datasets with several slides per patient.

METHODS

We collected multiple large multicentric datasets of colorectal cancer histopathological slides from the United Kingdom (FOXTROT, N = 21,384 slides; CR07, N = 7985 slides) and Germany (DACHS, N = 3606 slides). These datasets contained multiple types of tissue slides, including bowel resection specimens, endoscopic biopsies, lymph node resections, immunohistochemistry-stained slides, and tissue microarrays. We developed, trained, and tested a deep convolutional neural network model to predict the type of slide from the slide overview (thumbnail) image. The primary statistical endpoint was the macro-averaged area under the receiver operating curve (AUROCs) for detection of the type of slide.

RESULTS

In the primary dataset (FOXTROT), with an AUROC of 0.995 [95% confidence interval [CI]: 0.994-0.996] the algorithm achieved a high classification performance and was able to accurately predict the type of slide from the thumbnail image alone. In the two external test cohorts (CR07, DACHS) AUROCs of 0.982 [95% CI: 0.979-0.985] and 0.875 [95% CI: 0.864-0.887] were observed, which indicates the generalizability of the trained model on unseen datasets. With a confidence threshold of 0.95, the model reached an accuracy of 94.6% (7331 classified cases) in CR07 and 85.1% (2752 classified cases) for the DACHS cohort.

CONCLUSION

Our findings show that using the low-resolution thumbnail image is sufficient to accurately classify the type of slide in digital pathology. This can support researchers to make the vast resource of existing pathology archives accessible to modern AI models with only minimal manual annotations.

Prediction of DNA methylation-based tumor types from histopathology in central nervous system tumors with deep learning

Precision in the diagnosis of diverse central nervous system (CNS) tumor types is crucial for optimal treatment. DNA methylation profiles, which capture the methylation status of thousands of individual CpG sites, are state-of-the-art data-driven means to enhance diagnostic accuracy but are also time consuming and not widely available. Here, to address these limitations, we developed Deep lEarning from histoPathoLOgy and methYlation (DEPLOY), a deep learning model that classifies CNS tumors to ten major categories from histopathology. DEPLOY integrates three distinct components: the first classifies CNS tumors directly from slide images ('direct model'), the second initially generates predictions for DNA methylation beta values, which are subsequently used for tumor classification ('indirect model'), and the third classifies tumor types directly from routinely available patient demographics. First, we find that DEPLOY accurately predicts beta values from histopathology images. Second, using a ten-class model trained on an internal dataset of 1,796 patients, we predict the tumor categories in three independent external test datasets including 2,156 patients, achieving an overall accuracy of 95% and balanced accuracy of 91% on samples that are predicted with high confidence. These results showcase the potential future use of DEPLOY to assist pathologists in diagnosing CNS tumors within a clinically relevant short time frame.

Social network analysis of cell networks improves deep learning for prediction of molecular pathways and key mutations in colorectal cancer

Colorectal cancer (CRC) is a primary global health concern, and identifying the molecular pathways, genetic subtypes, and mutations associated with CRC is crucial for precision medicine. However, traditional measurement techniques such as gene sequencing are costly and time-consuming, while most deep learning methods proposed for this task lack interpretability. This study offers a new approach to enhance the state-of-the-art deep learning methods for molecular pathways and key mutation prediction by incorporating cell network information. We build cell graphs with nuclei as nodes and nuclei connections as edges of the network and leverage Social Network Analysis (SNA) measures to extract abstract, perceivable, and interpretable features that explicitly describe the cell network characteristics in an image. Our approach does not rely on precise nuclei segmentation or feature extraction, is computationally efficient, and is easily scalable. In this study, we utilize the TCGA-CRC-DX dataset, comprising 499 patients and 502 diagnostic slides from primary colorectal tumours, sourced from 36 distinct medical centres in the United States. By incorporating the SNA features alongside deep features in two multiple instance learning frameworks, we demonstrate improved performance for chromosomal instability (CIN), hypermutated tumour (HM), TP53 gene, BRAF gene, and Microsatellite instability (MSI) status prediction tasks (2.4%-4% and 7-8.8% improvement in AUROC and AUPRC on average). Additionally, our method achieves outstanding performance on MSI prediction in an external PAIP dataset (99% AUROC and 98% AUPRC), demonstrating its generalizability. Our findings highlight the discrimination power of SNA features and how they can be beneficial to deep learning models' performance and provide insights into the correlation of cell network profiles with molecular pathways and key mutations.

Immune subtyping of melanoma whole slide images using multiple instance learning

Determining early-stage prognostic markers and stratifying patients for effective treatment are two key challenges for improving outcomes for melanoma patients. Previous studies have used tumour transcriptome data to stratify patients into immune subgroups, which were associated with differential melanoma specific survival and potential predictive biomarkers. However, acquiring transcriptome data is a time-consuming and costly process. Moreover, it is not routinely used in the current clinical workflow. Here, we attempt to overcome this by developing deep learning models to classify gigapixel haematoxylin and eosin (H&E) stained pathology slides, which are well established in clinical workflows, into these immune subgroups. We systematically assess six different multiple instance learning (MIL) frameworks, using five different image resolutions and three different feature extraction methods. We show that pathology-specific self-supervised models using 10x resolution patches generate superior representations for the classification of immune subtypes. In addition, in a primary melanoma dataset, we achieve a mean area under the receiver operating characteristic curve (AUC) of 0.80 for classifying histopathology images into 'high' or 'low immune' subgroups and a mean AUC of 0.82 in an independent TCGA melanoma dataset. Furthermore, we show that these models are able to stratify patients into 'high' and 'low immune' subgroups with significantly different melanoma specific survival outcomes (log rank test, P< 0.005). We anticipate that MIL methods will allow us to find new biomarkers of high importance, act as a tool for clinicians to infer the immune landscape of tumours and stratify patients, without needing to carry out additional expensive genetic tests.

A systematic pan-cancer study on deep learning-based prediction of multi-omic biomarkers from routine pathology images

BACKGROUND

The objective of this comprehensive pan-cancer study is to evaluate the potential of deep learning (DL) for molecular profiling of multi-omic biomarkers directly from hematoxylin and eosin (H&E)-stained whole slide images.

METHODS

A total of 12,093 DL models predicting 4031 multi-omic biomarkers across 32 cancer types were trained and validated. The study included a broad range of genetic, transcriptomic, and proteomic biomarkers, as well as established prognostic markers, molecular subtypes, and clinical outcomes.

RESULTS

Here we show that 50% of the models achieve an area under the curve (AUC) of 0.644 or higher. The observed AUC for 25% of the models is at least 0.719 and exceeds 0.834 for the top 5%. Molecular profiling with image-based histomorphological features is generally considered feasible for most of the investigated biomarkers and across different cancer types. The performance appears to be independent of tumor purity, sample size, and class ratio (prevalence), suggesting a degree of inherent predictability in histomorphology.

CONCLUSIONS

The results demonstrate that DL holds promise to predict a wide range of biomarkers across the omics spectrum using only H&E-stained histological slides of solid tumors. This paves the way for accelerating diagnosis and developing more precise treatments for cancer patients.

A deep-learning workflow to predict upper tract urothelial carcinoma protein-based subtypes from H&E slides supporting the prioritization of patients for molecular testing

Upper tract urothelial carcinoma (UTUC) is a rare and aggressive, yet understudied, urothelial carcinoma (UC). The more frequent UC of the bladder comprises several molecular subtypes, associated with different targeted therapies and overlapping with protein-based subtypes. However, if and how these findings extend to UTUC remains unclear. Artificial intelligence-based approaches could help elucidate UTUC's biology and extend access to targeted treatments to a wider patient audience. Here, UTUC protein-based subtypes were identified, and a deep-learning (DL) workflow was developed to predict them directly from routine histopathological H&E slides. Protein-based subtypes in a retrospective cohort of 163 invasive tumors were assigned by hierarchical clustering of the immunohistochemical expression of three luminal (FOXA1, GATA3, and CK20) and three basal (CD44, CK5, and CK14) markers. Cluster analysis identified distinctive luminal (N = 80) and basal (N = 42) subtypes. The luminal subtype mostly included pushing, papillary tumors, whereas the basal subtype diffusely infiltrating, non-papillary tumors. DL model building relied on a transfer-learning approach by fine-tuning a pre-trained ResNet50. Classification performance was measured via three-fold repeated cross-validation. A mean area under the receiver operating characteristic curve of 0.83 (95% CI: 0.67-0.99), 0.8 (95% CI: 0.62-0.99), and 0.81 (95% CI: 0.65-0.96) was reached in the three repetitions. High-confidence DL-based predicted subtypes showed significant associations (p < 0.001) with morphological features, i.e. tumor type, histological subtypes, and infiltration type. Furthermore, a significant association was found with programmed cell death ligand 1 (PD-L1) combined positive score (p < 0.001) and FGFR3 mutational status (p = 0.002), with high-confidence basal predictions containing a higher proportion of PD-L1 positive samples and high-confidence luminal predictions a higher proportion of FGFR3-mutated samples. Testing of the DL model on an independent cohort highlighted the importance to accommodate histological subtypes. Taken together, our DL workflow can predict protein-based UTUC subtypes, associated with the presence of targetable alterations, directly from H&E slides.

Primary liver cancer classification from routine tumour biopsy using weakly supervised deep learning

Background & Aims

The diagnosis of primary liver cancers (PLCs) can be challenging, especially on biopsies and for combined hepatocellular-cholangiocarcinoma (cHCC-CCA). We automatically classified PLCs on routine-stained biopsies using a weakly supervised learning method.

Method

We selected 166 PLC biopsies divided into training, internal and external validation sets: 90, 29 and 47 samples, respectively. Two liver pathologists reviewed each whole-slide hematein eosin saffron (HES)-stained image (WSI). After annotating the tumour/non-tumour areas, tiles of 256x256 pixels were extracted from the WSIs and used to train a ResNet18 neural network. The tumour/non-tumour annotations served as labels during training, and the network's last convolutional layer was used to extract new tumour tile features. Without knowledge of the precise labels of the malignancies, we then applied an unsupervised clustering algorithm.

Results

Pathological review classified the training and validation sets into hepatocellular carcinoma (HCC, 33/90, 11/29 and 26/47), intrahepatic cholangiocarcinoma (iCCA, 28/90, 9/29 and 15/47), and cHCC-CCA (29/90, 9/29 and 6/47). In the two-cluster model, Clusters 0 and 1 contained mainly HCC and iCCA histological features. The diagnostic agreement between the pathological diagnosis and the two-cluster model predictions (major contingent) in the internal and external validation sets was 100% (11/11) and 96% (25/26) for HCC and 78% (7/9) and 87% (13/15) for iCCA, respectively. For cHCC-CCA, we observed a highly variable proportion of tiles from each cluster (cluster 0: 5-97%; cluster 1: 2-94%).

Conclusion

Our method applied to PLC HES biopsy could identify specific morphological features of HCC and iCCA. Although no specific features of cHCC-CCA were recognized, assessing the proportion of HCC and iCCA tiles within a slide could facilitate the identification of cHCC-CCA.

Impact and implications

The diagnosis of primary liver cancers can be challenging, especially on biopsies and for combined hepatocellular-cholangiocarcinoma (cHCC-CCA). We automatically classified primary liver cancers on routine-stained biopsies using a weakly supervised learning method. Our model identified specific features of hepatocellular carcinoma and intrahepatic cholangiocarcinoma. Despite no specific features of cHCC-CCA being recognized, the identification of hepatocellular carcinoma and intrahepatic cholangiocarcinoma tiles within a slide could facilitate the diagnosis of primary liver cancers, and particularly cHCC-CCA.

Tissue Contamination Challenges the Credibility of Machine Learning Models in Real World Digital Pathology

Machine learning (ML) models are poised to transform surgical pathology practice. The most successful use attention mechanisms to examine whole slides, identify which areas of tissue are diagnostic, and use them to guide diagnosis. Tissue contaminants, such as floaters, represent unexpected tissue. Although human pathologists are extensively trained to consider and detect tissue contaminants, we examined their impact on ML models. We trained 4 whole-slide models. Three operate in placenta for the following functions: (1) detection of decidual arteriopathy, (2) estimation of gestational age, and (3) classification of macroscopic placental lesions. We also developed a model to detect prostate cancer in needle biopsies. We designed experiments wherein patches of contaminant tissue are randomly sampled from known slides and digitally added to patient slides and measured model performance. We measured the proportion of attention given to contaminants and examined the impact of contaminants in the t-distributed stochastic neighbor embedding feature space. Every model showed performance degradation in response to one or more tissue contaminants. Decidual arteriopathy detection--balanced accuracy decreased from 0.74 to 0.69 ± 0.01 with addition of 1 patch of prostate tissue for every 100 patches of placenta (1% contaminant). Bladder, added at 10% contaminant, raised the mean absolute error in estimating gestational age from 1.626 weeks to 2.371 ± 0.003 weeks. Blood, incorporated into placental sections, induced false-negative diagnoses of intervillous thrombi. Addition of bladder to prostate cancer needle biopsies induced false positives, a selection of high-attention patches, representing 0.033 mm2, and resulted in a 97% false-positive rate when added to needle biopsies. Contaminant patches received attention at or above the rate of the average patch of patient tissue. Tissue contaminants induce errors in modern ML models. The high level of attention given to contaminants indicates a failure to encode biological phenomena. Practitioners should move to quantify and ameliorate this problem.

Extracting interpretable features for pathologists using weakly supervised learning to predict p16 expression in oropharyngeal cancer

One drawback of existing artificial intelligence (AI)-based histopathological prediction models is the lack of interpretability. The objective of this study is to extract p16-positive oropharyngeal squamous cell carcinoma (OPSCC) features in a form that can be interpreted by pathologists using AI model. We constructed a model for predicting p16 expression using a dataset of whole-slide images from 114 OPSCC biopsy cases. We used the clustering-constrained attention-based multiple-instance learning (CLAM) model, a weakly supervised learning approach. To improve performance, we incorporated tumor annotation into the model (Annot-CLAM) and achieved the mean area under the receiver operating characteristic curve of 0.905. Utilizing the image patches on which the model focused, we examined the features of model interest via histopathologic morphological analysis and cycle-consistent adversarial network (CycleGAN) image translation. The histopathologic morphological analysis evaluated the histopathological characteristics of image patches, revealing significant differences in the numbers of nuclei, the perimeters of the nuclei, and the intercellular bridges between p16-negative and p16-positive image patches. By using the CycleGAN-converted images, we confirmed that the sizes and densities of nuclei are significantly converted. This novel approach improves interpretability in histopathological morphology-based AI models and contributes to the advancement of clinically valuable histopathological morphological features.

Surrogate Biomarker Prediction from Whole-Slide Images for Evaluating Overall Survival in Lung Adenocarcinoma

BACKGROUND

Recent advances in computational pathology have shown potential in predicting biomarkers from haematoxylin and eosin (H&E) whole-slide images (WSI). However, predicting the outcome directly from WSIs remains a substantial challenge. In this study, we aimed to investigate how gene expression, predicted from WSIs, could be used to evaluate overall survival (OS) in patients with lung adenocarcinoma (LUAD).

METHODS

Differentially expressed genes (DEGs) were identified from The Cancer Genome Atlas (TCGA)-LUAD cohort. Cox regression analysis was performed on DEGs to identify the gene prognostics of OS. Attention-based multiple instance learning (AMIL) models were trained to predict the expression of identified prognostic genes from WSIs using the TCGA-LUAD dataset. Models were externally validated in the Clinical Proteomic Tumour Analysis Consortium (CPTAC)-LUAD dataset. The prognostic value of predicted gene expression values was then compared to the true gene expression measurements.

RESULTS

The expression of 239 prognostic genes could be predicted in TCGA-LUAD with cross-validated Pearson's R > 0.4. Predicted gene expression demonstrated prognostic performance, attaining a cross-validated concordance index of up to 0.615 in TCGA-LUAD through Cox regression. In total, 36 genes had predicted expression in the external validation cohort that was prognostic of OS.

CONCLUSIONS

Gene expression predicted from WSIs is an effective method of evaluating OS in patients with LUAD. These results may open up new avenues of cost- and time-efficient prognosis assessment in LUAD treatment.

Regression-based Deep-Learning predicts molecular biomarkers from pathology slides

Deep Learning (DL) can predict biomarkers from cancer histopathology. Several clinically approved applications use this technology. Most approaches, however, predict categorical labels, whereas biomarkers are often continuous measurements. We hypothesize that regression-based DL outperforms classification-based DL. Therefore, we develop and evaluate a self-supervised attention-based weakly supervised regression method that predicts continuous biomarkers directly from 11,671 images of patients across nine cancer types. We test our method for multiple clinically and biologically relevant biomarkers: homologous recombination deficiency score, a clinically used pan-cancer biomarker, as well as markers of key biological processes in the tumor microenvironment. Using regression significantly enhances the accuracy of biomarker prediction, while also improving the predictions' correspondence to regions of known clinical relevance over classification. In a large cohort of colorectal cancer patients, regression-based prediction scores provide a higher prognostic value than classification-based scores. Our open-source regression approach offers a promising alternative for continuous biomarker analysis in computational pathology.

Artificial Intelligence in Pathomics and Genomics of Renal Cell Carcinoma

The integration of artificial intelligence (AI) with histopathology images and gene expression patterns has led to the emergence of the dynamic fields of pathomics and genomics. These fields have revolutionized renal cell carcinoma (RCC) diagnosis and subtyping and improved survival prediction models. Machine learning has identified unique gene patterns across RCC subtypes and grades, providing insights into RCC origins and potential treatments, as targeted therapies. The combination of pathomics and genomics using AI opens new avenues in RCC research, promising future breakthroughs and innovations that patients and physicians can anticipate.

Encrypted federated learning for secure decentralized collaboration in cancer image analysis

Artificial intelligence (AI) has a multitude of applications in cancer research and oncology. However, the training of AI systems is impeded by the limited availability of large datasets due to data protection requirements and other regulatory obstacles. Federated and swarm learning represent possible solutions to this problem by collaboratively training AI models while avoiding data transfer. However, in these decentralized methods, weight updates are still transferred to the aggregation server for merging the models. This leaves the possibility for a breach of data privacy, for example by model inversion or membership inference attacks by untrusted servers. Somewhat-homomorphically-encrypted federated learning (SHEFL) is a solution to this problem because only encrypted weights are transferred, and model updates are performed in the encrypted space. Here, we demonstrate the first successful implementation of SHEFL in a range of clinically relevant tasks in cancer image analysis on multicentric datasets in radiology and histopathology. We show that SHEFL enables the training of AI models which outperform locally trained models and perform on par with models which are centrally trained. In the future, SHEFL can enable multiple institutions to co-train AI models without forsaking data governance and without ever transmitting any decryptable data to untrusted servers.

Weakly Supervised Deep Learning Predicts Immunotherapy Response in Solid Tumors Based on PD-L1 Expression

Programmed death-ligand 1 (PD-L1) IHC is the most commonly used biomarker for immunotherapy response. However, quantification of PD-L1 status in pathology slides is challenging. Neither manual quantification nor a computer-based mimicking of manual readouts is perfectly reproducible, and the predictive performance of both approaches regarding immunotherapy response is limited. In this study, we developed a deep learning (DL) method to predict PD-L1 status directly from raw IHC image data, without explicit intermediary steps such as cell detection or pigment quantification. We trained the weakly supervised model on PD-L1-stained slides from the non-small cell lung cancer (NSCLC)-Memorial Sloan Kettering (MSK) cohort (N = 233) and validated it on the pan-cancer-Vall d'Hebron Institute of Oncology (VHIO) cohort (N = 108). We also investigated the performance of the model to predict response to immune checkpoint inhibitors (ICI) in terms of progression-free survival. In the pan-cancer-VHIO cohort, the performance was compared with tumor proportion score (TPS) and combined positive score (CPS). The DL model showed good performance in predicting PD-L1 expression (TPS ≥ 1%) in both NSCLC-MSK and pan-cancer-VHIO cohort (AUC 0.88 ± 0.06 and 0.80 ± 0.03, respectively). The predicted PD-L1 status showed an improved association with response to ICIs [HR: 1.5 (95% confidence interval: 1-2.3), P = 0.049] compared with TPS [HR: 1.4 (0.96-2.2), P = 0.082] and CPS [HR: 1.2 (0.79-1.9), P = 0.386]. Notably, our explainability analysis showed that the model does not just look at the amount of brown pigment in the IHC slides, but also considers morphologic factors such as lymphocyte conglomerates. Overall, end-to-end weakly supervised DL shows potential for improving patient stratification for cancer immunotherapy by analyzing PD-L1 IHC, holistically integrating morphology and PD-L1 staining intensity.

SIGNIFICANCE

The weakly supervised DL model to predict PD-L1 status from raw IHC data, integrating tumor staining intensity and morphology, enables enhanced patient stratification in cancer immunotherapy compared with traditional pathologist assessment.

AdvMIL: Adversarial multiple instance learning for the survival analysis on whole-slide images

The survival analysis on histological whole-slide images (WSIs) is one of the most important means to estimate patient prognosis. Although many weakly-supervised deep learning models have been developed for gigapixel WSIs, their potential is generally restricted by classical survival analysis rules and fully-supervised learning requirements. As a result, these models provide patients only with a completely-certain point estimation of time-to-event, and they could only learn from the labeled WSI data currently at a small scale. To tackle these problems, we propose a novel adversarial multiple instance learning (AdvMIL) framework. This framework is based on adversarial time-to-event modeling, and integrates the multiple instance learning (MIL) that is much necessary for WSI representation learning. It is a plug-and-play one, so that most existing MIL-based end-to-end methods can be easily upgraded by applying this framework, gaining the improved abilities of survival distribution estimation and semi-supervised learning. Our extensive experiments show that AdvMIL not only could often bring performance improvement to mainstream WSI survival analysis methods at a relatively low computational cost, but also enables these methods to effectively utilize unlabeled data via semi-supervised learning. Moreover, it is observed that AdvMIL could help improving the robustness of models against patch occlusion and two representative image noises. The proposed AdvMIL framework could promote the research of survival analysis in computational pathology with its novel adversarial MIL paradigm.

SAMPLER: unsupervised representations for rapid analysis of whole slide tissue images

BACKGROUND

Deep learning has revolutionized digital pathology, allowing automatic analysis of hematoxylin and eosin (H&E) stained whole slide images (WSIs) for diverse tasks. WSIs are broken into smaller images called tiles, and a neural network encodes each tile. Many recent works use supervised attention-based models to aggregate tile-level features into a slide-level representation, which is then used for downstream analysis. Training supervised attention-based models is computationally intensive, architecture optimization of the attention module is non-trivial, and labeled data are not always available. Therefore, we developed an unsupervised and fast approach called SAMPLER to generate slide-level representations.

METHODS

Slide-level representations of SAMPLER are generated by encoding the cumulative distribution functions of multiscale tile-level features. To assess effectiveness of SAMPLER, slide-level representations of breast carcinoma (BRCA), non-small cell lung carcinoma (NSCLC), and renal cell carcinoma (RCC) WSIs of The Cancer Genome Atlas (TCGA) were used to train separate classifiers distinguishing tumor subtypes in FFPE and frozen WSIs. In addition, BRCA and NSCLC classifiers were externally validated on frozen WSIs. Moreover, SAMPLER's attention maps identify regions of interest, which were evaluated by a pathologist. To determine time efficiency of SAMPLER, we compared runtime of SAMPLER with two attention-based models. SAMPLER concepts were used to improve the design of a context-aware multi-head attention model (context-MHA).

FINDINGS

SAMPLER-based classifiers were comparable to state-of-the-art attention deep learning models to distinguish subtypes of BRCA (AUC = 0.911 ± 0.029), NSCLC (AUC = 0.940 ± 0.018), and RCC (AUC = 0.987 ± 0.006) on FFPE WSIs (internal test sets). However, training SAMLER-based classifiers was >100 times faster. SAMPLER models successfully distinguished tumor subtypes on both internal and external test sets of frozen WSIs. Histopathological review confirmed that SAMPLER-identified high attention tiles contained subtype-specific morphological features. The improved context-MHA distinguished subtypes of BRCA and RCC (BRCA-AUC = 0.921 ± 0.027, RCC-AUC = 0.988 ± 0.010) with increased accuracy on internal test FFPE WSIs.

INTERPRETATION

Our unsupervised statistical approach is fast and effective for analyzing WSIs, with greatly improved scalability over attention-based deep learning methods. The high accuracy of SAMPLER-based classifiers and interpretable attention maps suggest that SAMPLER successfully encodes the distinct morphologies within WSIs and will be applicable to general histology image analysis problems.

FUNDING

This study was supported by the National Cancer Institute (Grant No. R01CA230031 and P30CA034196).

Operational greenhouse-gas emissions of deep learning in digital pathology: a modelling study

BACKGROUND

Deep learning is a promising way to improve health care. Image-processing medical disciplines, such as pathology, are expected to be transformed by deep learning. The first clinically applicable deep-learning diagnostic support tools are already available in cancer pathology, and their number is increasing. However, data on the environmental sustainability of these tools are scarce. We aimed to conduct an environmental-sustainability analysis of a theoretical implementation of deep learning in patient-care pathology.

METHODS

For this modelling study, we first assembled and calculated relevant data and parameters of a digital-pathology workflow. Data were breast and prostate specimens from the university clinic at the Institute of Pathology of the Rheinisch-Westfälische Technische Hochschule Aachen (Aachen, Germany), for which commercially available deep learning was already available. Only specimens collected between Jan 1 and Dec 31, 2019 were used, to omit potential biases due to the COVID-19 pandemic. Our final selection was based on 2 representative weeks outside holidays, covering different types of specimens. To calculate carbon dioxide (CO2) or CO2 equivalent (CO2 eq) emissions of deep learning in pathology, we gathered relevant data for exact numbers and sizes of whole-slide images (WSIs), which were generated by scanning histopathology samples of prostate and breast specimens. We also evaluated different data input scenarios (including all slide tiles, only tiles containing tissue, or only tiles containing regions of interest). To convert estimated energy consumption from kWh to CO2 eq, we used the internet protocol address of the computational server and the Electricity Maps database to obtain information on the sources of the local electricity grid (ie, renewable vs non-renewable), and estimated the number of trees and proportion of the local and world's forests needed to sequester the CO2 eq emissions. We calculated the computational requirements and CO2 eq emissions of 30 deep-learning models that varied in task and size. The first scenario represented the use of one commercially available deep-learning model for one task in one case (1-task), the second scenario considered two deep-learning models for two tasks per case (2-task), the third scenario represented a future, potentially automated workflow that could handle 7 tasks per case (7-task), and the fourth scenario represented the use of a single potential, large, computer-vision model that could conduct multiple tasks (multitask). We also compared the performance (ie, accuracy) and CO2 eq emissions of different deep-learning models for the classification of renal cell carcinoma on WSIs, also from Rheinisch-Westfälische Technische Hochschule Aachen. We also tested other approaches to reducing CO2 eq emissions, including model pruning and an alternative method for histopathology analysis (pathomics).

FINDINGS

The pathology database contained 35 552 specimens (237 179 slides), 6420 of which were prostate specimens (10 115 slides) and 11 801 of which were breast specimens (19 763 slides). We selected and subsequently digitised 140 slides from eight breast-cancer cases and 223 slides from five prostate-cancer cases. Applying large deep-learning models on all WSI tiles of prostate and breast pathology cases would result in yearly CO2 eq emissions of 7·65 metric tons (t; 95% CI 7·62-7·68) with the use of a single deep-learning model per case; yearly CO2 eq emissions were up to 100·56 t (100·21-100·99) with the use of seven deep-learning models per case. CO2 eq emissions for different deep-learning model scenarios, data inputs, and deep-learning model sizes for all slides varied from 3·61 t (3·59-3·63) to 2795·30 t (1177·51-6482·13. For the estimated number of overall pathology cases worldwide, the yearly CO2 eq emissions varied, reaching up to 16 megatons (Mt) of CO2 eq, requiring up to 86 590 km2 (0·22%) of world forest to sequester the CO2 eq emissions. Use of the 7-task scenario and small deep-learning models on slides containing tissue only could substantially reduce CO2 eq emissions worldwide by up to 141 times (0·1 Mt, 95% CI 0·1-0·1). Considering the local environment in Aachen, Germany, the maximum CO2 eq emission from the use of deep learning in digital pathology only would require 32·8% (95% CI 13·8-76·6) of the local forest to sequester the CO2 eq emissions. A single pathomics run on a tissue could provide information that was comparable to or even better than the output of multitask deep-learning models, but with 147 times reduced CO2 eq emissions.

INTERPRETATION

Our findings suggest that widespread use of deep learning in pathology might have considerable global-warming potential. The medical community, policy decision makers, and the public should be aware of this potential and encourage the use of CO2 eq emissions reduction strategies where possible.

FUNDING

German Research Foundation, European Research Council, German Federal Ministry of Education and Research, Health, Economic Affairs and Climate Action, and the Innovation Fund of the Federal Joint Committee.

Recent Advances in Bioimage Analysis Methods for Detecting Skeletal Deformities in Biomedical and Aquaculture Fish Species

Detecting skeletal or bone-related deformities in model and aquaculture fish is vital for numerous biomedical studies. In biomedical research, model fish with bone-related disorders are potential indicators of various chemically induced toxins in their environment or poor dietary conditions. In aquaculture, skeletal deformities are affecting fish health, and economic losses are incurred by fish farmers. This survey paper focuses on showcasing the cutting-edge image analysis tools and techniques based on artificial intelligence that are currently applied in the analysis of bone-related deformities in aquaculture and model fish. These methods and tools play a significant role in improving research by automating various aspects of the analysis. This paper also sheds light on some of the hurdles faced when dealing with high-content bioimages and explores potential solutions to overcome these challenges.

Deep learning-based phenotyping reclassifies combined hepatocellular-cholangiocarcinoma

Primary liver cancer arises either from hepatocytic or biliary lineage cells, giving rise to hepatocellular carcinoma (HCC) or intrahepatic cholangiocarcinoma (ICCA). Combined hepatocellular- cholangiocarcinomas (cHCC-CCA) exhibit equivocal or mixed features of both, causing diagnostic uncertainty and difficulty in determining proper management. Here, we perform a comprehensive deep learning-based phenotyping of multiple cohorts of patients. We show that deep learning can reproduce the diagnosis of HCC vs. CCA with a high performance. We analyze a series of 405 cHCC-CCA patients and demonstrate that the model can reclassify the tumors as HCC or ICCA, and that the predictions are consistent with clinical outcomes, genetic alterations and in situ spatial gene expression profiling. This type of approach could improve treatment decisions and ultimately clinical outcome for patients with rare and biphenotypic cancers such as cHCC-CCA.

A Novel and Efficient Digital Pathology Classifier for Predicting Cancer Biomarkers Using Sequencer Architecture

In digital pathology tasks, transformers have achieved state-of-the-art results, surpassing convolutional neural networks (CNNs). However, transformers are usually complex and resource intensive. This study developed a novel and efficient digital pathology classifier called DPSeq to predict cancer biomarkers through fine-tuning a sequencer architecture integrating horizontal and vertical bidirectional long short-term memory networks. Using hematoxylin and eosin-stained histopathologic images of colorectal cancer from two international data sets (The Cancer Genome Atlas and Molecular and Cellular Oncology), the predictive performance of DPSeq was evaluated in a series of experiments. DPSeq demonstrated exceptional performance for predicting key biomarkers in colorectal cancer (microsatellite instability status, hypermutation, CpG island methylator phenotype status, BRAF mutation, TP53 mutation, and chromosomal instability), outperforming most published state-of-the-art classifiers in a within-cohort internal validation and a cross-cohort external validation. In addition, under the same experimental conditions using the same set of training and testing data sets, DPSeq surpassed four CNNs (ResNet18, ResNet50, MobileNetV2, and EfficientNet) and two transformer (Vision Transformer and Swin Transformer) models, achieving the highest area under the receiver operating characteristic curve and area under the precision-recall curve values in predicting microsatellite instability status, BRAF mutation, and CpG island methylator phenotype status. Furthermore, DPSeq required less time for both training and prediction because of its simple architecture. Therefore, DPSeq appears to be the preferred choice over transformer and CNN models for predicting cancer biomarkers.

Prediction models for hormone receptor status in female breast cancer do not extend to males: further evidence of sex-based disparity in breast cancer

Breast cancer prognosis and management for both men and women are reliant upon estrogen receptor alpha (ERα) and progesterone receptor (PR) expression to inform therapy. Previous studies have shown that there are sex-specific binding characteristics of ERα and PR in breast cancer and, counterintuitively, ERα expression is more common in male than female breast cancer. We hypothesized that these differences could have morphological manifestations that are undetectable to human observers but could be elucidated computationally. To investigate this, we trained attention-based multiple instance learning prediction models for ERα and PR using H&E-stained images of female breast cancer from the Cancer Genome Atlas (TCGA) (n = 1085) and deployed them on external female (n = 192) and male breast cancer images (n = 245). Both targets were predicted in the internal (AUROC for ERα prediction: 0.86 ± 0.02, p < 0.001; AUROC for PR prediction = 0.76 ± 0.03, p < 0.001) and external female cohorts (AUROC for ERα prediction: 0.78 ± 0.03, p < 0.001; AUROC for PR prediction = 0.80 ± 0.04, p < 0.001) but not the male cohort (AUROC for ERα prediction: 0.66 ± 0.14, p = 0.43; AUROC for PR prediction = 0.63 ± 0.04, p = 0.05). This suggests that subtle morphological differences invisible upon visual inspection may exist between the sexes, supporting previous immunohistochemical, genomic, and transcriptomic analyses.

Deep learning trained on lymph node status predicts outcome from gastric cancer histopathology: a retrospective multicentric study

AIM

Gastric cancer (GC) is a tumour entity with highly variant outcomes. Lymph node metastasis is a prognostically adverse biomarker. We hypothesised that GC primary tissue contains information that is predictive of lymph node status and patient prognosis and that this information can be extracted using deep learning (DL).

METHODS

Using three patient cohorts comprising 1146 patients, we trained and validated a DL system to predict lymph node status directly from haematoxylin and eosin-stained GC tissue sections. We investigated the concordance between the DL-based prediction from the primary tumour slides (aiN score) and the histopathological lymph node status (pN). Furthermore, we assessed the prognostic value of the aiN score alone and when combined with the pN status.

RESULTS

The aiN score predicted the pN status reaching area under the receiver operating characteristic curves of 0.71 in the training cohort and 0.69 and 0.65 in the two test cohorts. In a multivariate Cox analysis, the aiN score was an independent predictor of patient survival with hazard ratios of 1.5 in the training cohort and of 1.3 and 2.2 in the two test cohorts. A combination of the aiN score and the pN status prognostically stratified patients by survival with p-values <0.05 in logrank tests.

CONCLUSION

GC primary tumour tissue contains additional prognostic information that is accessible using the aiN score. In combination with the pN status, this can be used for personalised management of GC patients after prospective validation.

Neuropathologist-level integrated classification of adult-type diffuse gliomas using deep learning from whole-slide pathological images

Current diagnosis of glioma types requires combining both histological features and molecular characteristics, which is an expensive and time-consuming procedure. Determining the tumor types directly from whole-slide images (WSIs) is of great value for glioma diagnosis. This study presents an integrated diagnosis model for automatic classification of diffuse gliomas from annotation-free standard WSIs. Our model is developed on a training cohort (n = 1362) and a validation cohort (n = 340), and tested on an internal testing cohort (n = 289) and two external cohorts (n = 305 and 328, respectively). The model can learn imaging features containing both pathological morphology and underlying biological clues to achieve the integrated diagnosis. Our model achieves high performance with area under receiver operator curve all above 0.90 in classifying major tumor types, in identifying tumor grades within type, and especially in distinguishing tumor genotypes with shared histological features. This integrated diagnosis model has the potential to be used in clinical scenarios for automated and unbiased classification of adult-type diffuse gliomas.

Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study

Deep learning (DL) can accelerate the prediction of prognostic biomarkers from routine pathology slides in colorectal cancer (CRC). However, current approaches rely on convolutional neural networks (CNNs) and have mostly been validated on small patient cohorts. Here, we develop a new transformer-based pipeline for end-to-end biomarker prediction from pathology slides by combining a pre-trained transformer encoder with a transformer network for patch aggregation. Our transformer-based approach substantially improves the performance, generalizability, data efficiency, and interpretability as compared with current state-of-the-art algorithms. After training and evaluating on a large multicenter cohort of over 13,000 patients from 16 colorectal cancer cohorts, we achieve a sensitivity of 0.99 with a negative predictive value of over 0.99 for prediction of microsatellite instability (MSI) on surgical resection specimens. We demonstrate that resection specimen-only training reaches clinical-grade performance on endoscopic biopsy tissue, solving a long-standing diagnostic problem.

Generalizability of Self-Supervised Training Models for Digital Pathology: A Multicountry Comparison in Colorectal Cancer

PURPOSE

In this multicountry study, we aim to explore the effectiveness of self-supervised learning (SSL) in colorectal cancer (CRC)-related predictive tasks using large amount of unlabeled digital pathology imaging data.

METHODS

We adopted SimSiam to conduct self-supervised pretraining on two large whole-slide image CRC data sets from the United States and Australia. The SSL pretrained encoder is then used in several predictive tasks, including supervised predictive tasks (tissue classification, microsatellite instability v microsatellite stability classification), and weakly supervised predictive tasks (polyp type classification and adenoma grading, and 5-year survival prediction). Performance on the tasks was compared between models using SSL pretraining and those using ImageNet pretraining, and performance for one-country pretraining was compared with two-country pretraining.

RESULTS

We demonstrate that SSL pretraining outperforms ImageNet pretraining in predictive tasks, that is, SSL pretraining outperforms the ImageNet pretraining by 3.01% of F 1 score on average over supervised predictive tasks and 1.53% of AUC on average over weakly supervised predictive tasks. Furthermore, two-country SSL pretraining has shown more stable performance than single-country pretraining, that is, two-country pretraining outperforms at least one of the single-country pretrainings by 1.93% of F 1 on average over supervised predictive tasks and 1.36% of AUC on average over weakly-supervised predictive tasks.

CONCLUSION

We find that using unlabeled image data for SSL pretraining in CRC related tasks is more effective than using ImageNet pretraining. Furthermore, SSL pretraining using data from multiple countries achieve more stable performance and better generalization than single-country pretraining.

Whole Slide Imaging-Based Prediction of TP53 Mutations Identifies an Aggressive Disease Phenotype in Prostate Cancer

In prostate cancer, there is an urgent need for objective prognostic biomarkers that identify the metastatic potential of a tumor at an early stage. While recent analyses indicated TP53 mutations as candidate biomarkers, molecular profiling in a clinical setting is complicated by tumor heterogeneity. Deep learning models that predict the spatial presence of TP53 mutations in whole slide images (WSI) offer the potential to mitigate this issue. To assess the potential of WSIs as proxies for spatially resolved profiling and as biomarkers for aggressive disease, we developed TiDo, a deep learning model that achieves state-of-the-art performance in predicting TP53 mutations from WSIs of primary prostate tumors. In an independent multifocal cohort, the model showed successful generalization at both the patient and lesion level. Analysis of model predictions revealed that false positive (FP) predictions could at least partially be explained by TP53 deletions, suggesting that some FP carry an alteration that leads to the same histological phenotype as TP53 mutations. Comparative expression and histologic cell type analyses identified a TP53-like cellular phenotype triggered by expression of pathways affecting stromal composition. Together, these findings indicate that WSI-based models might not be able to perfectly predict the spatial presence of individual TP53 mutations but they have the potential to elucidate the prognosis of a tumor by depicting a downstream phenotype associated with aggressive disease biomarkers.

SIGNIFICANCE

Deep learning models predicting TP53 mutations from whole slide images of prostate cancer capture histologic phenotypes associated with stromal composition, lymph node metastasis, and biochemical recurrence, indicating their potential as in silico prognostic biomarkers. See related commentary by Bordeleau, p. 2809.

ChampKit: A framework for rapid evaluation of deep neural networks for patch-based histopathology classification

BACKGROUND AND OBJECTIVE

Histopathology is the gold standard for diagnosis of many cancers. Recent advances in computer vision, specifically deep learning, have facilitated the analysis of histopathology images for many tasks, including the detection of immune cells and microsatellite instability. However, it remains difficult to identify optimal models and training configurations for different histopathology classification tasks due to the abundance of available architectures and the lack of systematic evaluations. Our objective in this work is to present a software tool that addresses this need and enables robust, systematic evaluation of neural network models for patch classification in histology in a light-weight, easy-to-use package for both algorithm developers and biomedical researchers.

METHODS

Here we present ChampKit (Comprehensive Histopathology Assessment of Model Predictions toolKit): an extensible, fully reproducible evaluation toolkit that is a one-stop-shop to train and evaluate deep neural networks for patch classification. ChampKit curates a broad range of public datasets. It enables training and evaluation of models supported by timm directly from the command line, without the need for users to write any code. External models are enabled through a straightforward API and minimal coding. As a result, Champkit facilitates the evaluation of existing and new models and deep learning architectures on pathology datasets, making it more accessible to the broader scientific community. To demonstrate the utility of ChampKit, we establish baseline performance for a subset of possible models that could be employed with ChampKit, focusing on several popular deep learning models, namely ResNet18, ResNet50, and R26-ViT, a hybrid vision transformer. In addition, we compare each model trained either from random weight initialization or with transfer learning from ImageNet pretrained models. For ResNet18, we also consider transfer learning from a self-supervised pretrained model.

RESULTS

The main result of this paper is the ChampKit software. Using ChampKit, we were able to systemically evaluate multiple neural networks across six datasets. We observed mixed results when evaluating the benefits of pretraining versus random intialization, with no clear benefit except in the low data regime, where transfer learning was found to be beneficial. Surprisingly, we found that transfer learning from self-supervised weights rarely improved performance, which is counter to other areas of computer vision.

CONCLUSIONS

Choosing the right model for a given digital pathology dataset is nontrivial. ChampKit provides a valuable tool to fill this gap by enabling the evaluation of hundreds of existing (or user-defined) deep learning models across a variety of pathology tasks. Source code and data for the tool are freely accessible at https://github.com/SBU-BMI/champkit.

Deep learning-based subtyping of gastric cancer histology predicts clinical outcome: a multi-institutional retrospective study

INTRODUCTION

The Laurén classification is widely used for Gastric Cancer (GC) histology subtyping. However, this classification is prone to interobserver variability and its prognostic value remains controversial. Deep Learning (DL)-based assessment of hematoxylin and eosin (H&E) stained slides is a potentially useful tool to provide an additional layer of clinically relevant information, but has not been systematically assessed in GC.

OBJECTIVE

We aimed to train, test and externally validate a deep learning-based classifier for GC histology subtyping using routine H&E stained tissue sections from gastric adenocarcinomas and to assess its potential prognostic utility.

METHODS

We trained a binary classifier on intestinal and diffuse type GC whole slide images for a subset of the TCGA cohort (N = 166) using attention-based multiple instance learning. The ground truth of 166 GC was obtained by two expert pathologists. We deployed the model on two external GC patient cohorts, one from Europe (N = 322) and one from Japan (N = 243). We assessed classification performance using the Area Under the Receiver Operating Characteristic Curve (AUROC) and prognostic value (overall, cancer specific and disease free survival) of the DL-based classifier with uni- and multivariate Cox proportional hazard models and Kaplan-Meier curves with log-rank test statistics.

RESULTS

Internal validation using the TCGA GC cohort using five-fold cross-validation achieved a mean AUROC of 0.93 ± 0.07. External validation showed that the DL-based classifier can better stratify GC patients' 5-year survival compared to pathologist-based Laurén classification for all survival endpoints, despite frequently divergent model-pathologist classifications. Univariate overall survival Hazard Ratios (HRs) of pathologist-based Laurén classification (diffuse type versus intestinal type) were 1.14 (95% Confidence Interval (CI) 0.66-1.44, p-value = 0.51) and 1.23 (95% CI 0.96-1.43, p-value = 0.09) in the Japanese and European cohorts, respectively. DL-based histology classification resulted in HR of 1.46 (95% CI 1.18-1.65, p-value < 0.005) and 1.41 (95% CI 1.20-1.57, p-value < 0.005), in the Japanese and European cohorts, respectively. In diffuse type GC (as defined by the pathologist), classifying patients using the DL diffuse and intestinal classifications provided a superior survival stratification, and demonstrated statistically significant survival stratification when combined with pathologist classification for both the Asian (overall survival log-rank test p-value < 0.005, HR 1.43 (95% CI 1.05-1.66, p-value = 0.03) and European cohorts (overall survival log-rank test p-value < 0.005, HR 1.56 (95% CI 1.16-1.76, p-value < 0.005)).

CONCLUSION

Our study shows that gastric adenocarcinoma subtyping using pathologist's Laurén classification as ground truth can be performed using current state of the art DL techniques. Patient survival stratification seems to be better by DL-based histology typing compared with expert pathologist histology typing. DL-based GC histology typing has potential as an aid in subtyping. Further investigations are warranted to fully understand the underlying biological mechanisms for the improved survival stratification despite apparent imperfect classification by the DL algorithm.

A Weakly Supervised Deep Learning Framework for Whole Slide Classification to Facilitate Digital Pathology in Animal Study

The pathology of animal studies is crucial for toxicity evaluations and regulatory assessments, but the manual examination of slides by pathologists remains time-consuming and requires extensive training. One inherent challenge in this process is the interobserver variability, which can compromise the consistency and accuracy of a study. Artificial intelligence (AI) has demonstrated its ability to automate similar examinations in clinical applications with enhanced efficiency, consistency, and accuracy. However, training AI models typically relies on costly pixel-level annotation of injured regions and is often not available for animal pathology. To address this, we developed the PathologAI system, a "weakly" supervised approach for WSI classification in rat images without explicit lesion annotation at the pixel level. Using rat liver imaging data from the Open TG-GATEs system, PathologAI was applied to predict necrosis of n = 816 WSIs (377 controls). TG-GATEs studied 170 compounds at three dose levels (low, middle, and high) for four time points (3, 7, 14, and 28 days). PathologAI first preprocessed WSIs at the tile level to generate a high-level representation with a Generative Adversarial Network architecture. The prediction of liver necrosis relied on an ensemble model of 5 CNN classifiers trained on 335 WSIs. The ensemble model achieved notable classification accuracy on the holdout test set: 87% among 87 control slides free of findings, 83% among 120 controls with spontaneous necrosis, 67% among 147 treated animals with spontaneous minimal or slight necrosis, and 59% among 127 treated animals with minimal or slight necrosis caused by the treatment. Importantly, PathologAI was able to discriminate WSIs with spontaneous necrosis from those with treatment related necrosis and discriminated mild lesion level findings (slight vs minimal) and between treatment dose levels. PathologAI could provide an inexpensive and rapid screening tool to assist the digital pathology analysis in preclinical applications and general toxicological studies.

Machine learning for cryosection pathology predicts the 2021 WHO classification of glioma

BACKGROUND

Timely and accurate intraoperative cryosection evaluations remain the gold standard for guiding surgical treatments for gliomas. However, the tissue-freezing process often generates artifacts that make histologic interpretation difficult. In addition, the 2021 WHO Classification of Tumors of the Central Nervous System incorporates molecular profiles in the diagnostic categories, so standard visual evaluation of cryosections alone cannot completely inform diagnoses based on the new classification system.

METHODS

To address these challenges, we develop the context-aware Cryosection Histopathology Assessment and Review Machine (CHARM) using samples from 1,524 glioma patients from three different patient populations to systematically analyze cryosection slides.

FINDINGS

Our CHARM models successfully identified malignant cells (AUROC = 0.98 ± 0.01 in the independent validation cohort), distinguished isocitrate dehydrogenase (IDH)-mutant tumors from wild type (AUROC = 0.79-0.82), classified three major types of molecularly defined gliomas (AUROC = 0.88-0.93), and identified the most prevalent subtypes of IDH-mutant tumors (AUROC = 0.89-0.97). CHARM further predicts clinically important genetic alterations in low-grade glioma, including ATRX, TP53, and CIC mutations, CDKN2A/B homozygous deletion, and 1p/19q codeletion via cryosection images.

CONCLUSIONS

Our approaches accommodate the evolving diagnostic criteria informed by molecular studies, provide real-time clinical decision support, and will democratize accurate cryosection diagnoses.

FUNDING

Supported in part by the National Institute of General Medical Sciences grant R35GM142879, the Google Research Scholar Award, the Blavatnik Center for Computational Biomedicine Award, the Partners' Innovation Discovery Grant, and the Schlager Family Award for Early Stage Digital Health Innovations.

SAMPLER: Empirical distribution representations for rapid analysis of whole slide tissue images

Deep learning has revolutionized digital pathology, allowing for automatic analysis of hematoxylin and eosin (H&E) stained whole slide images (WSIs) for diverse tasks. In such analyses, WSIs are typically broken into smaller images called tiles, and a neural network backbone encodes each tile in a feature space. Many recent works have applied attention based deep learning models to aggregate tile-level features into a slide-level representation, which is then used for slide-level prediction tasks. However, training attention models is computationally intensive, necessitating hyperparameter optimization and specialized training procedures. Here, we propose SAMPLER, a fully statistical approach to generate efficient and informative WSI representations by encoding the empirical cumulative distribution functions (CDFs) of multiscale tile features. We demonstrate that SAMPLER-based classifiers are as accurate or better than state-of-the-art fully deep learning attention models for classification tasks including distinction of: subtypes of breast carcinoma (BRCA: AUC=0.911 ± 0.029); subtypes of non-small cell lung carcinoma (NSCLC: AUC=0.940±0.018); and subtypes of renal cell carcinoma (RCC: AUC=0.987±0.006). A major advantage of the SAMPLER representation is that predictive models are >100X faster compared to attention models. Histopathological review confirms that SAMPLER-identified high attention tiles contain tumor morphological features specific to the tumor type, while low attention tiles contain fibrous stroma, blood, or tissue folding artifacts. We further apply SAMPLER concepts to improve the design of attention-based neural networks, yielding a context aware multi-head attention model with increased accuracy for subtype classification within BRCA and RCC (BRCA: AUC=0.921±0.027, and RCC: AUC=0.988±0.010). Finally, we provide theoretical results identifying sufficient conditions for which SAMPLER is optimal. SAMPLER is a fast and effective approach for analyzing WSIs, with greatly improved scalability over attention methods to benefit digital pathology analysis.

Computational pathology in cancer diagnosis, prognosis, and prediction - present day and prospects

Computational pathology refers to applying deep learning techniques and algorithms to analyse and interpret histopathology images. Advances in artificial intelligence (AI) have led to an explosion in innovation in computational pathology, ranging from the prospect of automation of routine diagnostic tasks to the discovery of new prognostic and predictive biomarkers from tissue morphology. Despite the promising potential of computational pathology, its integration in clinical settings has been limited by a range of obstacles including operational, technical, regulatory, ethical, financial, and cultural challenges. Here, we focus on the pathologists' perspective of computational pathology: we map its current translational research landscape, evaluate its clinical utility, and address the more common challenges slowing clinical adoption and implementation. We conclude by describing contemporary approaches to drive forward these techniques. © 2023 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Unleashing the potential of AI for pathology: challenges and recommendations

Computational pathology is currently witnessing a surge in the development of AI techniques, offering promise for achieving breakthroughs and significantly impacting the practices of pathology and oncology. These AI methods bring with them the potential to revolutionize diagnostic pipelines as well as treatment planning and overall patient care. Numerous peer-reviewed studies reporting remarkable performance across diverse tasks serve as a testimony to the potential of AI in the field. However, widespread adoption of these methods in clinical and pre-clinical settings still remains a challenge. In this review article, we present a detailed analysis of the major obstacles encountered during the development of effective models and their deployment in practice. We aim to provide readers with an overview of the latest developments, assist them with insights into identifying some specific challenges that may require resolution, and suggest recommendations and potential future research directions. © 2023 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Recent advances of pathomics in colorectal cancer diagnosis and prognosis

Colorectal cancer (CRC) is one of the most common malignancies, with the third highest incidence and the second highest mortality in the world. To improve the therapeutic outcome, the risk stratification and prognosis predictions would help guide clinical treatment decisions. Achieving these goals have been facilitated by the fast development of artificial intelligence (AI) -based algorithms using radiological and pathological data, in combination with genomic information. Among them, features extracted from pathological images, termed pathomics, are able to reflect sub-visual characteristics linking to better stratification and prediction of therapeutic responses. In this paper, we review recent advances in pathological image-based algorithms in CRC, focusing on diagnosis of benign and malignant lesions, micro-satellite instability, as well as prediction of neoadjuvant chemoradiotherapy and the prognosis of CRC patients.

Gigapixel end-to-end training using streaming and attention

Current hardware limitations make it impossible to train convolutional neural networks on gigapixel image inputs directly. Recent developments in weakly supervised learning, such as attention-gated multiple instance learning, have shown promising results, but often use multi-stage or patch-wise training strategies risking suboptimal feature extraction, which can negatively impact performance. In this paper, we propose to train a ResNet-34 encoder with an attention-gated classification head in an end-to-end fashion, which we call StreamingCLAM, using a streaming implementation of convolutional layers. This allows us to train end-to-end on 4-gigapixel microscopic images using only slide-level labels. We achieve a mean area under the receiver operating characteristic curve of 0.9757 for metastatic breast cancer detection (CAMELYON16), close to fully supervised approaches using pixel-level annotations. Our model can also detect MYC-gene translocation in histologic slides of diffuse large B-cell lymphoma, achieving a mean area under the ROC curve of 0.8259. Furthermore, we show that our model offers a degree of interpretability through the attention mechanism.

A Weakly Supervised Deep Learning Model and Human-Machine Fusion for Accurate Grading of Renal Cell Carcinoma from Histopathology Slides

(1) Background: The Fuhrman grading (FG) system is widely used in the management of clear cell renal cell carcinoma (ccRCC). However, it is affected by observer variability and irreproducibility in clinical practice. We aimed to use a deep learning multi-class model called SSL-CLAM to assist in diagnosing the FG status of ccRCC patients using digitized whole slide images (WSIs). (2) Methods: We recruited 504 eligible ccRCC patients from The Cancer Genome Atlas (TCGA) cohort and obtained 708 hematoxylin and eosin-stained WSIs for the development and internal validation of the SSL-CLAM model. Additionally, we obtained 445 WSIs from 188 ccRCC eligible patients in the Clinical Proteomic Tumor Analysis Consortium (CPTAC) cohort as an independent external validation set. A human-machine fusion approach was used to validate the added value of the SSL-CLAM model for pathologists. (3) Results: The SSL-CLAM model successfully diagnosed the five FG statuses (Grade-0, 1, 2, 3, and 4) of ccRCC, and achieved AUCs of 0.917 and 0.887 on the internal and external validation sets, respectively, outperforming a junior pathologist. For the normal/tumor classification (Grade-0, Grade-1/2/3/4) task, the SSL-CLAM model yielded AUCs close to 1 on both the internal and external validation sets. The SSL-CLAM model achieved a better performance for the two-tiered FG (Grade-0, Grade-1/2, and Grade-3/4) task, with AUCs of 0.936 and 0.915 on the internal and external validation sets, respectively. The human-machine diagnostic performance was superior to that of the SSL-CLAM model, showing promising prospects. In addition, the high-attention regions of the SSL-CLAM model showed that with an increasing FG status, the cell nuclei in the tumor region become larger, with irregular contours and increased cellular pleomorphism. (4) Conclusions: Our findings support the feasibility of using deep learning and human-machine fusion methods for FG classification on WSIs from ccRCC patients, which may assist pathologists in making diagnostic decisions.

Semantic modeling of cell damage prediction: a machine learning approach at human-level performance in dermatology

Machine learning is transforming the field of histopathology. Especially in classification related tasks, there have been many successful applications of deep learning already. Yet, in tasks that rely on regression and many niche applications, the domain lacks cohesive procedures that are adapted to the learning processes of neural networks. In this work, we investigate cell damage in whole slide images of the epidermis. A common way for pathologists to annotate a score, characterizing the degree of damage for these samples, is the ratio between healthy and unhealthy nuclei. The annotation procedure of these scores, however, is expensive and prone to be noisy among pathologists. We propose a new measure of damage, that is the total area of damage, relative to the total area of the epidermis. In this work, we present results of regression and segmentation models, predicting both scores on a curated and public dataset. We have acquired the dataset in collaborative efforts with medical professionals. Our study resulted in a comprehensive evaluation of the proposed damage metrics in the epidermis, with recommendations, emphasizing practical relevance for real world applications.

Tissue contamination challenges the credibility of machine learning models in real world digital pathology

Machine learning (ML) models are poised to transform surgical pathology practice. The most successful use attention mechanisms to examine whole slides, identify which areas of tissue are diagnostic, and use them to guide diagnosis. Tissue contaminants, such as floaters, represent unexpected tissue. While human pathologists are extensively trained to consider and detect tissue contaminants, we examined their impact on ML models. We trained 4 whole slide models. Three operate in placenta for 1) detection of decidual arteriopathy (DA), 2) estimation of gestational age (GA), and 3) classification of macroscopic placental lesions. We also developed a model to detect prostate cancer in needle biopsies. We designed experiments wherein patches of contaminant tissue are randomly sampled from known slides and digitally added to patient slides and measured model performance. We measured the proportion of attention given to contaminants and examined the impact of contaminants in T-distributed Stochastic Neighbor Embedding (tSNE) feature space. Every model showed performance degradation in response to one or more tissue contaminants. DA detection balanced accuracy decreased from 0.74 to 0.69 +/- 0.01 with addition of 1 patch of prostate tissue for every 100 patches of placenta (1% contaminant). Bladder, added at 10% contaminant raised the mean absolute error in estimating gestation age from 1.626 weeks to 2.371 +/ 0.003 weeks. Blood, incorporated into placental sections, induced false negative diagnoses of intervillous thrombi. Addition of bladder to prostate cancer needle biopsies induced false positives, a selection of high-attention patches, representing 0.033mm2, resulted in a 97% false positive rate when added to needle biopsies. Contaminant patches received attention at or above the rate of the average patch of patient tissue. Tissue contaminants induce errors in modern ML models. The high level of attention given to contaminants indicates a failure to encode biological phenomena. Practitioners should move to quantify and ameliorate this problem.

Prediction of heart transplant rejection from routine pathology slides with self-supervised deep learning

Aims

One of the most important complications of heart transplantation is organ rejection, which is diagnosed on endomyocardial biopsies by pathologists. Computer-based systems could assist in the diagnostic process and potentially improve reproducibility. Here, we evaluated the feasibility of using deep learning in predicting the degree of cellular rejection from pathology slides as defined by the International Society for Heart and Lung Transplantation (ISHLT) grading system.

Methods and results

We collected 1079 histopathology slides from 325 patients from three transplant centres in Germany. We trained an attention-based deep neural network to predict rejection in the primary cohort and evaluated its performance using cross-validation and by deploying it to three cohorts. For binary prediction (rejection yes/no), the mean area under the receiver operating curve (AUROC) was 0.849 in the cross-validated experiment and 0.734, 0.729, and 0.716 in external validation cohorts. For a prediction of the ISHLT grade (0R, 1R, 2/3R), AUROCs were 0.835, 0.633, and 0.905 in the cross-validated experiment and 0.764, 0.597, and 0.913; 0.631, 0.633, and 0.682; and 0.722, 0.601, and 0.805 in the validation cohorts, respectively. The predictions of the artificial intelligence model were interpretable by human experts and highlighted plausible morphological patterns.

Conclusion

We conclude that artificial intelligence can detect patterns of cellular transplant rejection in routine pathology, even when trained on small cohorts.

Collaborative training of medical artificial intelligence models with non-uniform labels

Due to the rapid advancements in recent years, medical image analysis is largely dominated by deep learning (DL). However, building powerful and robust DL models requires training with large multi-party datasets. While multiple stakeholders have provided publicly available datasets, the ways in which these data are labeled vary widely. For Instance, an institution might provide a dataset of chest radiographs containing labels denoting the presence of pneumonia, while another institution might have a focus on determining the presence of metastases in the lung. Training a single AI model utilizing all these data is not feasible with conventional federated learning (FL). This prompts us to propose an extension to the widespread FL process, namely flexible federated learning (FFL) for collaborative training on such data. Using 695,000 chest radiographs from five institutions from across the globe-each with differing labels-we demonstrate that having heterogeneously labeled datasets, FFL-based training leads to significant performance increase compared to conventional FL training, where only the uniformly annotated images are utilized. We believe that our proposed algorithm could accelerate the process of bringing collaborative training methods from research and simulation phase to the real-world applications in healthcare.

Self-supervised attention-based deep learning for pan-cancer mutation prediction from histopathology

The histopathological phenotype of tumors reflects the underlying genetic makeup. Deep learning can predict genetic alterations from pathology slides, but it is unclear how well these predictions generalize to external datasets. We performed a systematic study on Deep-Learning-based prediction of genetic alterations from histology, using two large datasets of multiple tumor types. We show that an analysis pipeline that integrates self-supervised feature extraction and attention-based multiple instance learning achieves a robust predictability and generalizability.

Weakly supervised deep learning to predict recurrence in low-grade endometrial cancer from multiplexed immunofluorescence images

Predicting recurrence in low-grade, early-stage endometrial cancer (EC) is both challenging and clinically relevant. We present a weakly-supervised deep learning framework, NaroNet, that can learn, without manual expert annotation, the complex tumor-immune interrelations at three levels: local phenotypes, cellular neighborhoods, and tissue areas. It uses multiplexed immunofluorescence for the simultaneous visualization and quantification of CD68 + macrophages, CD8 + T cells, FOXP3 + regulatory T cells, PD-L1/PD-1 protein expression, and tumor cells. We used 489 tumor cores from 250 patients to train a multilevel deep-learning model to predict tumor recurrence. Using a tenfold cross-validation strategy, our model achieved an area under the curve of 0.90 with a 95% confidence interval of 0.83-0.95. Our model predictions resulted in concordance for 96,8% of cases (κ = 0.88). This method could accurately assess the risk of recurrence in EC, outperforming current prognostic factors, including molecular subtyping.

Generalizable biomarker prediction from cancer pathology slides with self-supervised deep learning: A retrospective multi-centric study

Deep learning (DL) can predict microsatellite instability (MSI) from routine histopathology slides of colorectal cancer (CRC). However, it is unclear whether DL can also predict other biomarkers with high performance and whether DL predictions generalize to external patient populations. Here, we acquire CRC tissue samples from two large multi-centric studies. We systematically compare six different state-of-the-art DL architectures to predict biomarkers from pathology slides, including MSI and mutations in BRAF, KRAS, NRAS, and PIK3CA. Using a large external validation cohort to provide a realistic evaluation setting, we show that models using self-supervised, attention-based multiple-instance learning consistently outperform previous approaches while offering explainable visualizations of the indicative regions and morphologies. While the prediction of MSI and BRAF mutations reaches a clinical-grade performance, mutation prediction of PIK3CA, KRAS, and NRAS was clinically insufficient.

A deep learning-based histopathology classifier for Focal Cortical Dysplasia

A light microscopy-based histopathology diagnosis of human brain specimens obtained from epilepsy surgery remains the gold standard to confirm the underlying cause of a patient’s focal epilepsy and further inform postsurgical patient management. The differential diagnosis of neocortical specimens in the realm of epilepsy surgery remains, however, challenging. Herein, we developed an open access, deep learning-based classifier to histopathologically assess whole slide microscopy images (WSI) and to automatically recognize various subtypes of Focal Cortical Dysplasia (FCD), according to the ILAE consensus classification update of 2022. We trained a convolutional neuronal network (CNN) with fully digitalized WSI of hematoxylin–eosin stainings obtained from 125 patients covering the spectrum of mild malformation of cortical development (mMCD), mMCD with oligodendroglial hyperplasia in epilepsy (MOGHE), FCD ILAE Type 1a, 2a and 2b using 414 formalin-fixed and paraffin-embedded archival tissue blocks. An additional series of 198 postmortem tissue blocks from 59 patients without neurological disorders served as control to train the CNN for homotypic frontal, temporal and occipital areas and heterotypic Brodmann areas 4 and 17, entorhinal cortex and dentate gyrus. Special stains and immunohistochemical reactions were used to comprehensively annotate the region of interest. We then programmed a novel tile extraction pipeline and graphical dashboard to visualize all areas on the WSI recognized by the CNN. Our deep learning-based classifier is able to compute 1000 × 1000 µm large tiles and recognizes 25 anatomical regions and FCD categories with an accuracy of 98.8% (F1 score = 0.82). Microscopic review of regions predicted by the network confirmed these results. This deep learning-based classifier will be made available as online web application to support the differential histopathology diagnosis in neocortical human brain specimens obtained from epilepsy surgery. It will also serve as blueprint to build a digital histopathology slide suite addressing all major brain diseases encountered in patients with surgically amenable focal epilepsy.

Predicting gene mutation status via artificial intelligence technologies based on multimodal integration (MMI) to advance precision oncology

Personalized treatment strategies for cancer frequently rely on the detection of genetic alterations which are determined by molecular biology assays. Historically, these processes typically required single-gene sequencing, next-generation sequencing, or visual inspection of histopathology slides by experienced pathologists in a clinical context. In the past decade, advances in artificial intelligence (AI) technologies have demonstrated remarkable potential in assisting physicians with accurate diagnosis of oncology image-recognition tasks. Meanwhile, AI techniques make it possible to integrate multimodal data such as radiology, histology, and genomics, providing critical guidance for the stratification of patients in the context of precision therapy. Given that the mutation detection is unaffordable and time-consuming for a considerable number of patients, predicting gene mutations based on routine clinical radiological scans or whole-slide images of tissue with AI-based methods has become a hot issue in actual clinical practice. In this review, we synthesized the general framework of multimodal integration (MMI) for molecular intelligent diagnostics beyond standard techniques. Then we summarized the emerging applications of AI in the prediction of mutational and molecular profiles of common cancers (lung, brain, breast, and other tumor types) pertaining to radiology and histology imaging. Furthermore, we concluded that there truly exist multiple challenges of AI techniques in the way of its real-world application in the medical field, including data curation, feature fusion, model interpretability, and practice regulations. Despite these challenges, we still prospect the clinical implementation of AI as a highly potential decision-support tool to aid oncologists in future cancer treatment management.

Radio-pathomic approaches in pediatric neuro-oncology: Opportunities and challenges

With medical software platforms moving to cloud environments with scalable storage and computing, the translation of predictive artificial intelligence (AI) models to aid in clinical decision-making and facilitate personalized medicine for cancer patients is becoming a reality. Medical imaging, namely radiologic and histologic images, has immense analytical potential in neuro-oncology, and models utilizing integrated radiomic and pathomic data may yield a synergistic effect and provide a new modality for precision medicine. At the same time, the ability to harness multi-modal data is met with challenges in aggregating data across medical departments and institutions, as well as significant complexity in modeling the phenotypic and genotypic heterogeneity of pediatric brain tumors. In this paper, we review recent pathomic and integrated pathomic, radiomic, and genomic studies with clinical applications. We discuss current challenges limiting translational research on pediatric brain tumors and outline technical and analytical solutions. Overall, we propose that to empower the potential residing in radio-pathomics, systemic changes in cross-discipline data management and end-to-end software platforms to handle multi-modal data sets are needed, in addition to embracing modern AI-powered approaches. These changes can improve the performance of predictive models, and ultimately the ability to advance brain cancer treatments and patient outcomes through the development of such models.

Deep Learning-Based Prediction of Molecular Tumor Biomarkers from H&E: A Practical Review

Molecular and genomic properties are critical in selecting cancer treatments to target individual tumors, particularly for immunotherapy. However, the methods to assess such properties are expensive, time-consuming, and often not routinely performed. Applying machine learning to H&E images can provide a more cost-effective screening method. Dozens of studies over the last few years have demonstrated that a variety of molecular biomarkers can be predicted from H&E alone using the advancements of deep learning: molecular alterations, genomic subtypes, protein biomarkers, and even the presence of viruses. This article reviews the diverse applications across cancer types and the methodology to train and validate these models on whole slide images. From bottom-up to pathologist-driven to hybrid approaches, the leading trends include a variety of weakly supervised deep learning-based approaches, as well as mechanisms for training strongly supervised models in select situations. While results of these algorithms look promising, some challenges still persist, including small training sets, rigorous validation, and model explainability. Biomarker prediction models may yield a screening method to determine when to run molecular tests or an alternative when molecular tests are not possible. They also create new opportunities in quantifying intratumoral heterogeneity and predicting patient outcomes.

Deep learning radiomics under multimodality explore association between muscle/fat and metastasis and survival in breast cancer patients

Sarcopenia is correlated with poor clinical outcomes in breast cancer (BC) patients. However, there is no precise quantitative study on the correlation between body composition changes and BC metastasis and survival. The present study proposed a deep learning radiomics (DLR) approach to investigate the effects of muscle and fat on distant metastasis and death outcomes in BC patients. Image feature extraction was performed on 4th thoracic vertebra (T4) and 11th thoracic vertebra (T11) on computed tomography (CT) image levels by DLR, and image features were combined with clinical information to predict distant metastasis in BC patients. Clinical information combined with DLR significantly predicted distant metastasis in BC patients. In the test cohort, the area under the curve of model performance on clinical information combined with DLR was 0.960 (95% CI: 0.942-0.979, P < 0.001). The patients with distant metastases had a lower pectoral muscle index in T4 (PMI/T4) than in patients without metastases. PMI/T4 and visceral fat tissue area in T11 (VFA/T11) were independent prognostic factors for the overall survival in BC patients. The pectoralis muscle area in T4 (PMA/T4) and PMI/T4 is an independent prognostic factor for distant metastasis-free survival in BC patients. The current study further confirmed that muscle/fat of T4 and T11 levels have a significant effect on the distant metastasis of BC. Appending the network features of T4 and T11 to the model significantly enhances the prediction performance of distant metastasis of BC, providing a valuable biomarker for the early treatment of BC patients.