Layer-Wise Relevance Propagation for Explaining Deep Neural Network Decisions in MRI-Based Alzheimer's Disease Classification

Interpretable brain disease classification and relevance-guided deep learning

Deep neural networks are increasingly used for neurological disease classification by MRI, but the networks' decisions are not easily interpretable by humans. Heat mapping by deep Taylor decomposition revealed that (potentially misleading) image features even outside of the brain tissue are crucial for the classifier's decision. We propose a regularization technique to train convolutional neural network (CNN) classifiers utilizing relevance-guided heat maps calculated online during training. The method was applied using T1-weighted MR images from 128 subjects with Alzheimer's disease (mean age = 71.9 ± 8.5 years) and 290 control subjects (mean age = 71.3 ± 6.4 years). The developed relevance-guided framework achieves higher classification accuracies than conventional CNNs but more importantly, it relies on less but more relevant and physiological plausible voxels within brain tissue. Additionally, preprocessing effects from skull stripping and registration are mitigated. With the interpretability of the decision mechanisms underlying CNNs, these results challenge the notion that unprocessed T1-weighted brain MR images in standard CNNs yield higher classification accuracy in Alzheimer's disease than solely atrophy.

Classification of Alzheimer's Disease Based on Weakly Supervised Learning and Attention Mechanism

The brain lesions images of Alzheimer's disease (AD) patients are slightly different from the Magnetic Resonance Imaging of normal people, and the classification effect of general image recognition technology is not ideal. Alzheimer's datasets are small, making it difficult to train large-scale neural networks. In this paper, we propose a network model (WS-AMN) that fuses weak supervision and an attention mechanism. The weakly supervised data augmentation network is used as the basic model, the attention map generated by weakly supervised learning is used to guide the data augmentation, and an attention module with channel domain and spatial domain is embedded in the residual network to focus on the distinctive channels and spaces of images respectively. The location information enhances the corresponding features of related features and suppresses the influence of irrelevant features.The results show that the F1-score is 99.63%, the accuracy is 99.61%. Our model provides a high-performance solution for accurate classification of AD.

Advancing brain network models to reconcile functional neuroimaging and clinical research

Functional magnetic resonance imaging (fMRI) captures information on brain function beyond the anatomical alterations that are traditionally visually examined by neuroradiologists. However, the fMRI signals are complex in addition to being noisy, so fMRI still faces limitations for clinical applications. Here we review methods that have been proposed as potential solutions so far, namely statistical, biophysical and decoding models, with their strengths and weaknesses. We especially evaluate the ability of these models to directly predict clinical variables from their parameters (predictability) and to extract clinically relevant information regarding biological mechanisms and relevant features for classification and prediction (interpretability). We then provide guidelines for useful applications and pitfalls of such fMRI-based models in a clinical research context, looking beyond the current state of the art. In particular, we argue that the clinical relevance of fMRI calls for a new generation of models for fMRI data, which combine the strengths of both biophysical and decoding models. This leads to reliable and biologically meaningful model parameters, which thus fulfills the need for simultaneous interpretability and predictability. In our view, this synergy is fundamental for the discovery of new pharmacological and interventional targets, as well as the use of models as biomarkers in neurology and psychiatry.

Application of explainable artificial intelligence for healthcare: A systematic review of the last decade (2011-2022)

BACKGROUND AND OBJECTIVES

Artificial intelligence (AI) has branched out to various applications in healthcare, such as health services management, predictive medicine, clinical decision-making, and patient data and diagnostics. Although AI models have achieved human-like performance, their use is still limited because they are seen as a black box. This lack of trust remains the main reason for their low use in practice, especially in healthcare. Hence, explainable artificial intelligence (XAI) has been introduced as a technique that can provide confidence in the model's prediction by explaining how the prediction is derived, thereby encouraging the use of AI systems in healthcare. The primary goal of this review is to provide areas of healthcare that require more attention from the XAI research community.

METHODS

Multiple journal databases were thoroughly searched using PRISMA guidelines 2020. Studies that do not appear in Q1 journals, which are highly credible, were excluded.

RESULTS

In this review, we surveyed 99 Q1 articles covering the following XAI techniques: SHAP, LIME, GradCAM, LRP, Fuzzy classifier, EBM, CBR, rule-based systems, and others.

CONCLUSION

We discovered that detecting abnormalities in 1D biosignals and identifying key text in clinical notes are areas that require more attention from the XAI research community. We hope this is review will encourage the development of a holistic cloud system for a smart city.

Development of a deep learning network for Alzheimer’s disease classification with evaluation of imaging modality and longitudinal data

Objective. Neuroimaging uncovers important information about disease in the brain. Yet in Alzheimer’s disease (AD), there remains a clear clinical need for reliable tools to extract diagnoses from neuroimages. Significant work has been done to develop deep learning (DL) networks using neuroimaging for AD diagnosis. However, no particular model has emerged as optimal. Due to a lack of direct comparisons and evaluations on independent data, there is no consensus on which modality is best for diagnostic models or whether longitudinal information enhances performance. The purpose of this work was (1) to develop a generalizable DL model to distinguish neuroimaging scans of AD patients from controls and (2) to evaluate the influence of imaging modality and longitudinal data on performance. Approach. We trained a 2-class convolutional neural network (CNN) with and without a cascaded recurrent neural network (RNN). We used datasets of 772 (N AD = 364, N control = 408) 3D 18F-FDG PET scans and 780 (N AD = 280, N control = 500) T1-weighted volumetric-3D MR images (containing 131 and 144 patients with multiple timepoints) from the Alzheimer’s Disease Neuroimaging Initiative, plus an independent set of 104 (N AD = 63, N NC = 41) 18F-FDG PET scans (one per patient) for validation. Main Results. ROC analysis showed that PET-trained models outperformed MRI-trained, achieving maximum AUC with the CNN + RNN model of 0.93 ± 0.08, with accuracy 82.5 ± 8.9%. Adding longitudinal information offered significant improvement to performance on 18F-FDG PET, but not on T1-MRI. CNN model validation with an independent 18F-FDG PET dataset achieved AUC of 0.99. Layer-wise relevance propagation heatmaps added CNN interpretability. Significance. The development of a high-performing tool for AD diagnosis, with the direct evaluation of key influences, reveals the advantage of using 18F-FDG PET and longitudinal data over MRI and single timepoint analysis. This has significant implications for the potential of neuroimaging for future research on AD diagnosis and clinical management of suspected AD patients.

Interpretable deep learning-based hippocampal sclerosis classification

OBJECTIVE

To evaluate the performance of a deep learning model for hippocampal sclerosis classification on the clinical dataset and suggest plausible visual interpretation for the model prediction.

METHODS

T2-weighted oblique coronal images of the brain MRI epilepsy protocol performed on patients were used. The training set included 320 participants with 160 no, 100 left and 60 right hippocampal sclerosis, and cross-validation was implemented. The test set consisted of 302 participants with 252 no, 25 left and 25 right hippocampal sclerosis. As the test set was imbalanced, we took an average of the accuracy achieved within each group to measure a balanced accuracy for multiclass and binary classifications. The dataset was composed to include not only healthy participants but also participants with abnormalities besides hippocampal sclerosis in the control group. We visualized the reasons for the model prediction using the layer-wise relevance propagation method.

RESULTS

When evaluated on the validation of the training set, we achieved multiclass and binary classification accuracy of 87.5% and 88.8% from the voting ensemble of six models. Evaluated on the test sets, we achieved multiclass and binary classification accuracy of 91.5% and 89.76%. The distinctly sparse visual interpretations were provided for each individual participant and group to suggest the contribution of each input voxel to the prediction on the MRI.

SIGNIFICANCE

The current interpretable deep learning-based model is promising for adapting effectively to clinical settings by utilizing commonly used data, such as MRI, with realistic abnormalities faced by neurologists to support the diagnosis of hippocampal sclerosis with plausible visual interpretation.

Using Occlusion-Based Saliency Maps to Explain an Artificial Intelligence Tool in Lung Cancer Screening: Agreement Between Radiologists, Labels, and Visual Prompts

Occlusion-based saliency maps (OBSMs) are one of the approaches for interpreting decision-making process of an artificial intelligence (AI) system. This study explores the agreement among text responses from a cohort of radiologists to describe diagnostically relevant areas on low-dose CT (LDCT) images. It also explores if radiologists' descriptions of cases misclassified by the AI provide a rationale for ruling out the AI's output. The OBSM indicating the importance of different pixels on the final decision made by an AI were generated for 10 benign cases (3 misclassified by the AI tool as malignant) and 10 malignant cases (2 misclassified by the AI tool as benign). Thirty-six radiologists were asked to use radiological vocabulary, typical to reporting LDCT scans, to describe the mapped regions of interest (ROI). The radiologists' annotations were then grouped by using a clustering-based technique. Topics were extracted from the annotations and for each ROI, a percentage of annotations containing each topic were found. Radiologists annotated 17 and 24 unique ROIs on benign and malignant cases, respectively. Agreement on the main label (e.g., "vessel," "nodule") by radiologists was only seen in only in 12% of all areas (5/41 ROI). Topic analyses identified six descriptors which are commonly associated with a lower malignancy likelihood. Eight common topics related to a higher malignancy likelihood were also determined. Occlusion-based saliency maps were used to explain an AI decision-making process to radiologists, who in turn have provided insight into the level of agreement between the AI's decision and radiological lexicon.

Explainable, trustworthy, and ethical machine learning for healthcare: A survey

With the advent of machine learning (ML) and deep learning (DL) empowered applications for critical applications like healthcare, the questions about liability, trust, and interpretability of their outputs are raising. The black-box nature of various DL models is a roadblock to clinical utilization. Therefore, to gain the trust of clinicians and patients, we need to provide explanations about the decisions of models. With the promise of enhancing the trust and transparency of black-box models, researchers are in the phase of maturing the field of eXplainable ML (XML). In this paper, we provided a comprehensive review of explainable and interpretable ML techniques for various healthcare applications. Along with highlighting security, safety, and robustness challenges that hinder the trustworthiness of ML, we also discussed the ethical issues arising because of the use of ML/DL for healthcare. We also describe how explainable and trustworthy ML can resolve all these ethical problems. Finally, we elaborate on the limitations of existing approaches and highlight various open research problems that require further development.

Identification of Unique Genetic Biomarkers of Various Subtypes of Glomerulonephritis Using Machine Learning and Deep Learning

(1) Objective: Identification of potential genetic biomarkers for various glomerulonephritis (GN) subtypes and discovering the molecular mechanisms of GN. (2) Methods: four microarray datasets of GN were downloaded from Gene Expression Omnibus (GEO) database and merged to obtain the gene expression profiles of eight GN subtypes. Then, differentially expressed immune-related genes (DIRGs) were identified to explore the molecular mechanisms of GN, and single-sample gene set enrichment analysis (ssGSEA) was performed to discover the abnormal inflammation in GN. In addition, a nomogram model was generated using the R package "glmnet", and the calibration curve was plotted to evaluate the predictive power of the nomogram model. Finally, deep learning (DL) based on a multilayer perceptron (MLP) network was performed to explore the characteristic genes for GN. (3) Results: we screened out 274 common up-regulated or down-regulated DIRGs in the glomeruli and tubulointerstitium. These DIRGs are mainly involved in T-cell differentiation, the RAS signaling pathway, and the MAPK signaling pathway. ssGSEA indicates that there is a significant increase in DC (dendritic cells) and macrophages, and a significant decrease in neutrophils and NKT cells in glomeruli, while monocytes and NK cells are increased in tubulointerstitium. A nomogram model was constructed to predict GN based on 7 DIRGs, and 20 DIRGs of each subtype of GN in glomeruli and tubulointerstitium were selected as characteristic genes. (4) Conclusions: this study reveals that the DIRGs are closely related to the pathogenesis of GN and could serve as genetic biomarkers in GN. DL further identified the characteristic genes that are essential to define the pathogenesis of GN and develop targeted therapies for eight GN subtypes.

Ensemble learning using traditional machine learning and deep neural network for diagnosis of Alzheimer's disease

In recent years, Alzheimer's disease (AD) diagnosis using neuroimaging and deep learning has drawn great research attention. However, due to the scarcity of training neuroimaging data, many deep learning models have suffered from severe overfitting. In this study, we propose an ensemble learning framework that combines deep learning and machine learning. The deep learning model was based on a 3D-ResNet to exploit 3D structural features of neuroimaging data. Meanwhile, Extreme Gradient Boosting (XGBoost) machine learning was applied on a voxel-wise basis to draw the most significant voxel groups out of the image. The 3D-ResNet and XGBoost predictions were combined with patient demographics and cognitive test scores (Mini-Mental State Examination (MMSE) and Clinical Dementia Rating (CDR)) to give a final diagnosis prediction. Our proposed method was trained and validated on brain MRI brain images of the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. During the training phase, multiple data augmentation methods were employed to tackle overfitting. Our test set contained only baseline scans, i.e., the first visit scans since we aimed to investigate the ability of our approach in detecting AD during the first visit of AD patients. Our 5-fold cross-validation implementation achieved an average AUC of 100% during training and 96% during testing. Using the same computer, our method was much faster in scoring a prediction, approximately 10 min, than feature extraction-based machine learning methods, which often take many hours to score a prediction. To make the prediction explainable, we visualized the brain MRI image regions that primarily affected the 3D-ResNet model's prediction via heatmap. Lastly, we observed that proper generation of test sets was critical to avoiding the data leakage issue and ensuring the validity of results.

Mitigating Bias in Radiology Machine Learning: 3. Performance Metrics

The increasing use of machine learning (ML) algorithms in clinical settings raises concerns about bias in ML models. Bias can arise at any step of ML creation, including data handling, model development, and performance evaluation. Potential biases in the ML model can be minimized by implementing these steps correctly. This report focuses on performance evaluation and discusses model fitness, as well as a set of performance evaluation toolboxes: namely, performance metrics, performance interpretation maps, and uncertainty quantification. By discussing the strengths and limitations of each toolbox, our report highlights strategies and considerations to mitigate and detect biases during performance evaluations of radiology artificial intelligence models. Keywords: Segmentation, Diagnosis, Convolutional Neural Network (CNN) © RSNA, 2022.

Mapping the association between tau-PET and Aβ-amyloid-PET using deep learning

In Alzheimer’s disease, the molecular pathogenesis of the extracellular Aβ-amyloid (Aβ) instigation of intracellular tau accumulation is poorly understood. We employed a high-resolution PET scanner, with low detection thresholds, to examine the Aβ-tau association using a convolutional neural network (CNN), and compared results to a standard voxel-wise linear analysis. The full range of Aβ Centiloid values was highly predicted by the tau topography using the CNN (training R² = 0.86, validation R² = 0.75, testing R² = 0.72). Linear models based on tau-SUVR identified widespread positive correlations between tau accumulation and Aβ burden throughout the brain. In contrast, CNN analysis identified focal clusters in the bilateral medial temporal lobes, frontal lobes, precuneus, postcentral gyrus and middle cingulate. At low Aβ levels, information from the middle cingulate, frontal lobe and precuneus regions was more predictive of Aβ burden, while at high Aβ levels, the medial temporal regions were more predictive of Aβ burden. The data-driven CNN approach revealed new associations between tau topography and Aβ burden.

A robust framework to investigate the reliability and stability of explainable artificial intelligence markers of Mild Cognitive Impairment and Alzheimer's Disease

In clinical practice, several standardized neuropsychological tests have been designed to assess and monitor the neurocognitive status of patients with neurodegenerative diseases such as Alzheimer's disease. Important research efforts have been devoted so far to the development of multivariate machine learning models that combine the different test indexes to predict the diagnosis and prognosis of cognitive decline with remarkable results. However, less attention has been devoted to the explainability of these models. In this work, we present a robust framework to (i) perform a threefold classification between healthy control subjects, individuals with cognitive impairment, and subjects with dementia using different cognitive indexes and (ii) analyze the variability of the explainability SHAP values associated with the decisions taken by the predictive models. We demonstrate that the SHAP values can accurately characterize how each index affects a patient's cognitive status. Furthermore, we show that a longitudinal analysis of SHAP values can provide effective information on Alzheimer's disease progression.

Towards the Interpretability of Deep Learning Models for Multi-modal Neuroimaging: Finding Structural Changes of the Ageing Brain

Brain-age (BA) estimates based on deep learning are increasingly used as neuroimaging biomarker for brain health; however, the underlying neural features have remained unclear. We combined ensembles of convolutional neural networks with Layer-wise Relevance Propagation (LRP) to detect which brain features contribute to BA. Trained on magnetic resonance imaging (MRI) data of a population-based study (n=2637, 18-82 years), our models estimated age accurately based on single and multiple modalities, regionally restricted and whole-brain images (mean absolute errors 3.37-3.86 years). We find that BA estimates capture aging at both small and large-scale changes, revealing gross enlargements of ventricles and subarachnoid spaces, as well as white matter lesions, and atrophies that appear throughout the brain. Divergence from expected aging reflected cardiovascular risk factors and accelerated aging was more pronounced in the frontal lobe. Applying LRP, our study demonstrates how superior deep learning models detect brain-aging in healthy and at-risk individuals throughout adulthood.

Interpreting Brain Biomarkers: Challenges and solutions in interpreting machine learning-based predictive neuroimaging

Predictive modeling of neuroimaging data (predictive neuroimaging) for evaluating individual differences in various behavioral phenotypes and clinical outcomes is of growing interest. However, the field is experiencing challenges regarding the interpretability of the results. Approaches to defining the specific contribution of functional connections, regions, or networks in prediction models are urgently needed, which may help explore the underlying mechanisms. In this article, we systematically review the methods and applications for interpreting brain signatures derived from predictive neuroimaging based on a survey of 326 research articles. Strengths, limitations, and the suitable conditions for major interpretation strategies are also deliberated. In-depth discussion of common issues in existing literature and the corresponding recommendations to address these pitfalls are provided. We highly recommend exhaustive validation on the reliability and interpretability of the biomarkers across multiple datasets and contexts, which thereby could translate technical advances in neuroimaging into concrete improvements in precision medicine.

A deep learning model to classify and detect brain abnormalities in portable microwave based imaging system

Automated classification and detection of brain abnormalities like a tumor(s) in reconstructed microwave (RMW) brain images are essential for medical application investigation and monitoring disease progression. This paper presents the automatic classification and detection of human brain abnormalities through the deep learning-based YOLOv5 object detection model in a portable microwave head imaging system (MWHI). Initially, four hundred RMW image samples, including non-tumor and tumor(s) in different locations are collected from the implemented MWHI system. The RMW image dimension is 640 × 640 pixels. After that, image pre-processing and augmentation techniques are applied to generate the training dataset, consisting of 4400 images. Later, 80% of images are used to train the models, and 20% are used for testing. Later, from the 80% training dataset, 20% are utilized to validate the models. The detection and classification performances are evaluated by three variations of the YOLOv5 model: YOLOv5s, YOLOv5m, and YOLOv5l. It is investigated that the YOLOv5l model performed better compared to YOLOv5s, YOLOv5m, and state-of-the-art object detection models. The achieved accuracy, precision, sensitivity, specificity, F1-score, mean average precision (mAP), and classification loss are 96.32%, 95.17%, 94.98%, 95.28%, 95.53%, 96.12%, and 0.0130, respectively for the YOLOv5l model. The YOLOv5l model automatically detected tumor(s) accurately with a predicted bounding box including objectness score in RMW images and classified the tumors into benign and malignant classes. So, the YOLOv5l object detection model can be reliable for automatic tumor(s) detection and classification in a portable microwave brain imaging system as a real-time application.

A Comprehensive Survey on the Detection, Classification, and Challenges of Neurological Disorders

Neurological disorders (NDs) are becoming more common, posing a concern to pregnant women, parents, healthy infants, and children. Neurological disorders arise in a wide variety of forms, each with its own set of origins, complications, and results. In recent years, the intricacy of brain functionalities has received a better understanding due to neuroimaging modalities, such as magnetic resonance imaging (MRI), magnetoencephalography (MEG), and positron emission tomography (PET), etc. With high-performance computational tools and various machine learning (ML) and deep learning (DL) methods, these modalities have discovered exciting possibilities for identifying and diagnosing neurological disorders. This study follows a computer-aided diagnosis methodology, leading to an overview of pre-processing and feature extraction techniques. The performance of existing ML and DL approaches for detecting NDs is critically reviewed and compared in this article. A comprehensive portion of this study also shows various modalities and disease-specified datasets that detect and records images, signals, and speeches, etc. Limited related works are also summarized on NDs, as this domain has significantly fewer works focused on disease and detection criteria. Some of the standard evaluation metrics are also presented in this study for better result analysis and comparison. This research has also been outlined in a consistent workflow. At the conclusion, a mandatory discussion section has been included to elaborate on open research challenges and directions for future work in this emerging field.

A Progressive and Cross-Domain Deep Transfer Learning Framework for Wrist Fracture Detection

There has been an amplified focus on and benefit from the adoption of artificial intelligence (AI) in medical imaging applications. However, deep learning approaches involve training with massive amounts of annotated data in order to guarantee generalization and achieve high accuracies. Gathering and annotating large sets of training images require expertise which is both expensive and time-consuming, especially in the medical field. Furthermore, in health care systems where mistakes can have catastrophic consequences, there is a general mistrust in the black-box aspect of AI models. In this work, we focus on improving the performance of medical imaging applications when limited data is available while focusing on the interpretability aspect of the proposed AI model. This is achieved by employing a novel transfer learning framework, progressive transfer learning, an automated annotation technique and a correlation analysis experiment on the learned representations.

Progressive transfer learning helps jump-start the training of deep neural networks while improving the performance by gradually transferring knowledge from two source tasks into the target task. It is empirically tested on the wrist fracture detection application by first training a general radiology network RadiNet and using its weights to initialize RadiNetwrist , that is trained on wrist images to detect fractures. Experiments show that RadiNetwrist achieves an accuracy of 87% and an AUC ROC of 94% as opposed to 83% and 92% when it is pre-trained on the ImageNet dataset.

This improvement in performance is investigated within an explainable AI framework. More concretely, the learned deep representations of RadiNetwrist are compared to those learned by the baseline model by conducting a correlation analysis experiment. The results show that, when transfer learning is gradually applied, some features are learned earlier in the network. Moreover, the deep layers in the progressive transfer learning framework are shown to encode features that are not encountered when traditional transfer learning techniques are applied.

In addition to the empirical results, a clinical study is conducted and the performance of RadiNetwrist is compared to that of an expert radiologist. We found that RadiNetwrist exhibited similar performance to that of radiologists with more than 20 years of experience.

This motivates follow-up research to train on more data to feasibly surpass radiologists’ performance, and investigate the interpretability of AI models in the healthcare domain where the decision-making process needs to be credible and transparent.

The added value of PSMA PET/MR radiomics for prostate cancer staging

PURPOSE

To evaluate the performance of combined PET and multiparametric MRI (mpMRI) radiomics for the group-wise prediction of postsurgical Gleason scores (psGSs) in primary prostate cancer (PCa) patients.

METHODS

Patients with PCa, who underwent [68 Ga]Ga-PSMA-11 PET/MRI followed by radical prostatectomy, were included in this retrospective analysis (n = 101). Patients were grouped by psGS in three categories: ISUP grades 1-3, ISUP grade 4, and ISUP grade 5. mpMRI images included T1-weighted, T2-weighted, and apparent diffusion coefficient (ADC) map. Whole-prostate segmentations were performed on each modality, and image biomarker standardization initiative (IBSI)-compliant radiomic features were extracted. Nine support vector machine (SVM) models were trained: four single-modality radiomic models (PET, T1w, T2w, ADC); three PET + MRI double-modality models (PET + T1w, PET + T2w, PET + ADC), and two baseline models (one with patient data, one image-based) for comparison. A sixfold stratified cross-validation was performed, and balanced accuracies (bAcc) of the predictions of the best-performing models were reported and compared through Student's t-tests. The predictions of the best-performing model were compared against biopsy GS (bGS).

RESULTS

All radiomic models outperformed the baseline models. The best-performing (mean ± stdv [%]) single-modality model was the ADC model (76 ± 6%), although not significantly better (p > 0.05) than other single-modality models (T1w: 72 ± 3%, T2w: 73 ± 2%; PET: 75 ± 5%). The overall best-performing model combined PET + ADC radiomics (82 ± 5%). It significantly outperformed most other double-modality (PET + T1w: 74 ± 5%, p = 0.026; PET + T2w: 71 ± 4%, p = 0.003) and single-modality models (PET: p = 0.042; T1w: p = 0.002; T2w: p = 0.003), except the ADC-only model (p = 0.138). In this initial cohort, the PET + ADC model outperformed bGS overall (82.5% vs 72.4%) in the prediction of psGS.

CONCLUSION

All single- and double-modality models outperformed the baseline models, showing their potential in the prediction of GS, even with an unbalanced cohort. The best-performing model included PET + ADC radiomics, suggesting a complementary value of PSMA-PET and ADC radiomics.

Patch individual filter layers in CNNs to harness the spatial homogeneity of neuroimaging data

Convolutional neural networks (CNNs)-as a type of deep learning-have been specifically designed for highly heterogeneous data, such as natural images. Neuroimaging data, however, is comparably homogeneous due to (1) the uniform structure of the brain and (2) additional efforts to spatially normalize the data to a standard template using linear and non-linear transformations. To harness spatial homogeneity of neuroimaging data, we suggest here a new CNN architecture that combines the idea of hierarchical abstraction in CNNs with a prior on the spatial homogeneity of neuroimaging data. Whereas early layers are trained globally using standard convolutional layers, we introduce patch individual filters (PIF) for higher, more abstract layers. By learning filters in individual latent space patches without sharing weights, PIF layers can learn abstract features faster and specific to regions. We thoroughly evaluated PIF layers for three different tasks and data sets, namely sex classification on UK Biobank data, Alzheimer's disease detection on ADNI data and multiple sclerosis detection on private hospital data, and compared it with two baseline models, a standard CNN and a patch-based CNN. We obtained two main results: First, CNNs using PIF layers converge consistently faster, measured in run time in seconds and number of iterations than both baseline models. Second, both the standard CNN and the PIF model outperformed the patch-based CNN in terms of balanced accuracy and receiver operating characteristic area under the curve (ROC AUC) with a maximal balanced accuracy (ROC AUC) of 94.21% (99.10%) for the sex classification task (PIF model), and 81.24% and 80.48% (88.89% and 87.35%) respectively for the Alzheimer's disease and multiple sclerosis detection tasks (standard CNN model). In conclusion, we demonstrated that CNNs using PIF layers result in faster convergence while obtaining the same predictive performance as a standard CNN. To the best of our knowledge, this is the first study that introduces a prior in form of an inductive bias to harness spatial homogeneity of neuroimaging data.

Modelling and identification of characteristic kinematic features preceding freezing of gait with convolutional neural networks and layer-wise relevance propagation

BACKGROUND

Although deep neural networks (DNNs) are showing state of the art performance in clinical gait analysis, they are considered to be black-box algorithms. In other words, there is a lack of direct understanding of a DNN's ability to identify relevant features, hindering clinical acceptance. Interpretability methods have been developed to ameliorate this concern by providing a way to explain DNN predictions.

METHODS

This paper proposes the use of an interpretability method to explain DNN decisions for classifying the movement that precedes freezing of gait (FOG), one of the most debilitating symptoms of Parkinson's disease (PD). The proposed two-stage pipeline consists of (1) a convolutional neural network (CNN) to model the reduction of movement present before a FOG episode, and (2) layer-wise relevance propagation (LRP) to visualize the underlying features that the CNN perceives as important to model the pathology. The CNN was trained with the sagittal plane kinematics from a motion capture dataset of fourteen PD patients with FOG. The robustness of the model predictions and learned features was further assessed on fourteen PD patients without FOG and fourteen age-matched healthy controls.

RESULTS

The CNN proved highly accurate in modelling the movement that precedes FOG, with 86.8% of the strides being correctly identified. However, the CNN model was unable to model the movement for one of the seven patients that froze during the protocol. The LRP interpretability case study shows that (1) the kinematic features perceived as most relevant by the CNN are the reduced peak knee flexion and the fixed ankle dorsiflexion during the swing phase, (2) very little relevance for FOG is observed in the PD patients without FOG and the healthy control subjects, and (3) the poor predictive performance of one subject is attributed to the patient's unique and severely flexed gait signature.

CONCLUSIONS

The proposed pipeline can aid clinicians in explaining DNN decisions in clinical gait analysis and aid machine learning practitioners in assessing the generalization of their models by ensuring that the predictions are based on meaningful kinematic features.

Data-driven identification of diagnostically useful extrastriatal signal in dopamine transporter SPECT using explainable AI

This study used explainable artificial intelligence for data-driven identification of extrastriatal brain regions that can contribute to the interpretation of dopamine transporter SPECT with ¹²³I-FP-CIT in parkinsonian syndromes. A total of 1306 ¹²³I-FP-CIT-SPECT were included retrospectively. Binary classification as ‘reduced’ or ‘normal’ striatal ¹²³I-FP-CIT uptake by an experienced reader served as standard-of-truth. A custom-made 3-dimensional convolutional neural network (CNN) was trained for classification of the SPECT images with 1006 randomly selected images in three different settings: “full image”, “striatum only” (3-dimensional region covering the striata cropped from the full image), “without striatum” (full image with striatal region removed). The remaining 300 SPECT images were used to test the CNN classification performance. Layer-wise relevance propagation (LRP) was used for voxelwise quantification of the relevance for the CNN-based classification in this test set. Overall accuracy of CNN-based classification was 97.0%, 95.7%, and 69.3% in the “full image”, “striatum only”, and “without striatum” setting. Prominent contributions in the LRP-based relevance maps beyond the striatal signal were detected in insula, amygdala, ventromedial prefrontal cortex, thalamus, anterior temporal cortex, superior frontal lobe, and pons, suggesting that ¹²³I-FP-CIT uptake in these brain regions provides clinically useful information for the differentiation of neurodegenerative and non-neurodegenerative parkinsonian syndromes.

Improving 3D convolutional neural network comprehensibility via interactive visualization of relevance maps: evaluation in Alzheimer's disease

Background

Although convolutional neural networks (CNNs) achieve high diagnostic accuracy for detecting Alzheimer’s disease (AD) dementia based on magnetic resonance imaging (MRI) scans, they are not yet applied in clinical routine. One important reason for this is a lack of model comprehensibility. Recently developed visualization methods for deriving CNN relevance maps may help to fill this gap as they allow the visualization of key input image features that drive the decision of the model. We investigated whether models with higher accuracy also rely more on discriminative brain regions predefined by prior knowledge.

Methods

We trained a CNN for the detection of AD in N = 663 T1-weighted MRI scans of patients with dementia and amnestic mild cognitive impairment (MCI) and verified the accuracy of the models via cross-validation and in three independent samples including in total N = 1655 cases. We evaluated the association of relevance scores and hippocampus volume to validate the clinical utility of this approach. To improve model comprehensibility, we implemented an interactive visualization of 3D CNN relevance maps, thereby allowing intuitive model inspection.

Results

Across the three independent datasets, group separation showed high accuracy for AD dementia versus controls (AUC ≥ 0.91) and moderate accuracy for amnestic MCI versus controls (AUC ≈ 0.74). Relevance maps indicated that hippocampal atrophy was considered the most informative factor for AD detection, with additional contributions from atrophy in other cortical and subcortical regions. Relevance scores within the hippocampus were highly correlated with hippocampal volumes (Pearson’s r ≈ −0.86, p < 0.001).

Conclusion

The relevance maps highlighted atrophy in regions that we had hypothesized a priori. This strengthens the comprehensibility of the CNN models, which were trained in a purely data-driven manner based on the scans and diagnosis labels. The high hippocampus relevance scores as well as the high performance achieved in independent samples support the validity of the CNN models in the detection of AD-related MRI abnormalities. The presented data-driven and hypothesis-free CNN modeling approach might provide a useful tool to automatically derive discriminative features for complex diagnostic tasks where clear clinical criteria are still missing, for instance for the differential diagnosis between various types of dementia.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-021-00924-2.

DeepAtrophy: Teaching a neural network to detect progressive changes in longitudinal MRI of the hippocampal region in Alzheimer's disease

Measures of change in hippocampal volume derived from longitudinal MRI are a well-studied biomarker of disease progression in Alzheimer's disease (AD) and are used in clinical trials to track therapeutic efficacy of disease-modifying treatments. However, longitudinal MRI change measures based on deformable registration can be confounded by MRI artifacts, resulting in over-estimation or underestimation of hippocampal atrophy. For example, the deformation-based-morphometry method ALOHA (Das et al., 2012) finds an increase in hippocampal volume in a substantial proportion of longitudinal scan pairs from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, unexpected, given that the hippocampal gray matter is lost with age and disease progression. We propose an alternative approach to quantify disease progression in the hippocampal region: to train a deep learning network (called DeepAtrophy) to infer temporal information from longitudinal scan pairs. The underlying assumption is that by learning to derive time-related information from scan pairs, the network implicitly learns to detect progressive changes that are related to aging and disease progression. Our network is trained using two categorical loss functions: one that measures the network's ability to correctly order two scans from the same subject, input in arbitrary order; and another that measures the ability to correctly infer the ratio of inter-scan intervals between two pairs of same-subject input scans. When applied to longitudinal MRI scan pairs from subjects unseen during training, DeepAtrophy achieves greater accuracy in scan temporal ordering and interscan interval inference tasks than ALOHA (88.5% vs. 75.5% and 81.1% vs. 75.0%, respectively). A scalar measure of time-related change in a subject level derived from DeepAtrophy is then examined as a biomarker of disease progression in the context of AD clinical trials. We find that this measure performs on par with ALOHA in discriminating groups of individuals at different stages of the AD continuum. Overall, our results suggest that using deep learning to infer temporal information from longitudinal MRI of the hippocampal region has good potential as a biomarker of disease progression, and hints that combining this approach with conventional deformation-based morphometry algorithms may lead to improved biomarkers in the future.

Explainable AI to improve acceptance of convolutional neural networks for automatic classification of dopamine transporter SPECT in the diagnosis of clinically uncertain parkinsonian syndromes

Purpose

Deep convolutional neural networks (CNN) provide high accuracy for automatic classification of dopamine transporter (DAT) SPECT images. However, CNN are inherently black-box in nature lacking any kind of explanation for their decisions. This limits their acceptance for clinical use. This study tested layer-wise relevance propagation (LRP) to explain CNN-based classification of DAT-SPECT in patients with clinically uncertain parkinsonian syndromes.

Methods

The study retrospectively included 1296 clinical DAT-SPECT with visual binary interpretation as “normal” or “reduced” by two experienced readers as standard-of-truth. A custom-made CNN was trained with 1008 randomly selected DAT-SPECT. The remaining 288 DAT-SPECT were used to assess classification performance of the CNN and to test LRP for explanation of the CNN-based classification.

Results

Overall accuracy, sensitivity, and specificity of the CNN were 95.8%, 92.8%, and 98.7%, respectively. LRP provided relevance maps that were easy to interpret in each individual DAT-SPECT. In particular, the putamen in the hemisphere most affected by nigrostriatal degeneration was the most relevant brain region for CNN-based classification in all reduced DAT-SPECT. Some misclassified DAT-SPECT showed an “inconsistent” relevance map more typical for the true class label.

Conclusion

LRP is useful to provide explanation of CNN-based decisions in individual DAT-SPECT and, therefore, can be recommended to support CNN-based classification of DAT-SPECT in clinical routine. Total computation time of 3 s is compatible with busy clinical workflow. The utility of “inconsistent” relevance maps to identify misclassified cases requires further investigation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00259-021-05569-9.

Risk Prediction of Cardiovascular Events by Exploration of Molecular Data with Explainable Artificial Intelligence

Cardiovascular diseases (CVD) annually take almost 18 million lives worldwide. Most lethal events occur months or years after the initial presentation. Indeed, many patients experience repeated complications or require multiple interventions (recurrent events). Apart from affecting the individual, this leads to high medical costs for society. Personalized treatment strategies aiming at prediction and prevention of recurrent events rely on early diagnosis and precise prognosis. Complementing the traditional environmental and clinical risk factors, multi-omics data provide a holistic view of the patient and disease progression, enabling studies to probe novel angles in risk stratification. Specifically, predictive molecular markers allow insights into regulatory networks, pathways, and mechanisms underlying disease. Moreover, artificial intelligence (AI) represents a powerful, yet adaptive, framework able to recognize complex patterns in large-scale clinical and molecular data with the potential to improve risk prediction. Here, we review the most recent advances in risk prediction of recurrent cardiovascular events, and discuss the value of molecular data and biomarkers for understanding patient risk in a systems biology context. Finally, we introduce explainable AI which may improve clinical decision systems by making predictions transparent to the medical practitioner.

Interpretable deep learning as a means for decrypting disease signature in multiple sclerosis

Objective.The mechanisms driving multiple sclerosis (MS) are still largely unknown, calling for new methods allowing to detect and characterize tissue degeneration since the early stages of the disease. Our aim is to decrypt the microstructural signatures of the Primary Progressive versus the Relapsing-Remitting state of disease based on diffusion and structural magnetic resonance imaging data.Approach.A selection of microstructural descriptors, based on the 3D-Simple Harmonics Oscillator Based Reconstruction and Estimation and the set of new algebraically independent Rotation Invariant spherical harmonics Features, was considered and used to feed convolutional neural networks (CNNs) models. Classical measures derived from diffusion tensor imaging, that are fractional anisotropy and mean diffusivity, were used as benchmark for diffusion MRI (dMRI). Finally, T1-weighted images were also considered for the sake of comparison with the state-of-the-art. A CNN model was fit to each feature map and layerwise relevance propagation (LRP) heatmaps were generated for each model, target class and subject in the test set. Average heatmaps were calculated across correctly classified patients and size-corrected metrics were derived on a set of regions of interest to assess the LRP contrast between the two classes.Main results.Our results demonstrated that dMRI features extracted in grey matter tissues can help in disambiguating primary progressive multiple sclerosis from relapsing-remitting multiple sclerosis patients and, moreover, that LRP heatmaps highlight areas of high relevance which relate well with what is known from literature for MS disease.Significance.Within a patient stratification task, LRP allows detecting the input voxels that mostly contribute to the classification of the patients in either of the two classes for each feature, potentially bringing to light hidden data properties which might reveal peculiar disease-state factors.

Interpretable deep learning for the remote characterisation of ambulation in multiple sclerosis using smartphones

The emergence of digital technologies such as smartphones in healthcare applications have demonstrated the possibility of developing rich, continuous, and objective measures of multiple sclerosis (MS) disability that can be administered remotely and out-of-clinic. Deep Convolutional Neural Networks (DCNN) may capture a richer representation of healthy and MS-related ambulatory characteristics from the raw smartphone-based inertial sensor data than standard feature-based methodologies. To overcome the typical limitations associated with remotely generated health data, such as low subject numbers, sparsity, and heterogeneous data, a transfer learning (TL) model from similar large open-source datasets was proposed. Our TL framework leveraged the ambulatory information learned on human activity recognition (HAR) tasks collected from wearable smartphone sensor data. It was demonstrated that fine-tuning TL DCNN HAR models towards MS disease recognition tasks outperformed previous Support Vector Machine (SVM) feature-based methods, as well as DCNN models trained end-to-end, by upwards of 8-15%. A lack of transparency of "black-box" deep networks remains one of the largest stumbling blocks to the wider acceptance of deep learning for clinical applications. Ensuing work therefore aimed to visualise DCNN decisions attributed by relevance heatmaps using Layer-Wise Relevance Propagation (LRP). Through the LRP framework, the patterns captured from smartphone-based inertial sensor data that were reflective of those who are healthy versus people with MS (PwMS) could begin to be established and understood. Interpretations suggested that cadence-based measures, gait speed, and ambulation-related signal perturbations were distinct characteristics that distinguished MS disability from healthy participants. Robust and interpretable outcomes, generated from high-frequency out-of-clinic assessments, could greatly augment the current in-clinic assessment picture for PwMS, to inform better disease management techniques, and enable the development of better therapeutic interventions.

Anthropogenic influence on extreme precipitation over global land areas seen in multiple observational datasets

The intensification of extreme precipitation under anthropogenic forcing is robustly projected by global climate models, but highly challenging to detect in the observational record. Large internal variability distorts this anthropogenic signal. Models produce diverse magnitudes of precipitation response to anthropogenic forcing, largely due to differing schemes for parameterizing subgrid-scale processes. Meanwhile, multiple global observational datasets of daily precipitation exist, developed using varying techniques and inhomogeneously sampled data in space and time. Previous attempts to detect human influence on extreme precipitation have not incorporated model uncertainty, and have been limited to specific regions and observational datasets. Using machine learning methods that can account for these uncertainties and capable of identifying the time evolution of the spatial patterns, we find a physically interpretable anthropogenic signal that is detectable in all global observational datasets. Machine learning efficiently generates multiple lines of evidence supporting detection of an anthropogenic signal in global extreme precipitation.

Which GAN? A comparative study of generative adversarial network-based fast MRI reconstruction

Fast magnetic resonance imaging (MRI) is crucial for clinical applications that can alleviate motion artefacts and increase patient throughput. K-space undersampling is an obvious approach to accelerate MR acquisition. However, undersampling of k-space data can result in blurring and aliasing artefacts for the reconstructed images. Recently, several studies have been proposed to use deep learning-based data-driven models for MRI reconstruction and have obtained promising results. However, the comparison of these methods remains limited because the models have not been trained on the same datasets and the validation strategies may be different. The purpose of this work is to conduct a comparative study to investigate the generative adversarial network (GAN)-based models for MRI reconstruction. We reimplemented and benchmarked four widely used GAN-based architectures including DAGAN, ReconGAN, RefineGAN and KIGAN. These four frameworks were trained and tested on brain, knee and liver MRI images using twofold, fourfold and sixfold accelerations, respectively, with a random undersampling mask. Both quantitative evaluations and qualitative visualization have shown that the RefineGAN method has achieved superior performance in reconstruction with better accuracy and perceptual quality compared to other GAN-based methods. This article is part of the theme issue 'Synergistic tomographic image reconstruction: part 1'.

Interpretation and visualization techniques for deep learning models in medical imaging

Deep learning (DL) approaches to medical image analysis tasks have recently become popular; however, they suffer from a lack of human interpretability critical for both increasing understanding of the methods' operation and enabling clinical translation. This review summarizes currently available methods for performing image model interpretation and critically evaluates published uses of these methods for medical imaging applications. We divide model interpretation in two categories: (1) understanding model structure and function and (2) understanding model output. Understanding model structure and function summarizes ways to inspect the learned features of the model and how those features act on an image. We discuss techniques for reducing the dimensionality of high-dimensional data and cover autoencoders, both of which can also be leveraged for model interpretation. Understanding model output covers attribution-based methods, such as saliency maps and class activation maps, which produce heatmaps describing the importance of different parts of an image to the model prediction. We describe the mathematics behind these methods, give examples of their use in medical imaging, and compare them against one another. We summarize several published toolkits for model interpretation specific to medical imaging applications, cover limitations of current model interpretation methods, provide recommendations for DL practitioners looking to incorporate model interpretation into their task, and offer general discussion on the importance of model interpretation in medical imaging contexts.

Promises and pitfalls of deep neural networks in neuroimaging-based psychiatric research

By promising more accurate diagnostics and individual treatment recommendations, deep neural networks and in particular convolutional neural networks have advanced to a powerful tool in medical imaging. Here, we first give an introduction into methodological key concepts and resulting methodological promises including representation and transfer learning, as well as modelling domain-specific priors. After reviewing recent applications within neuroimaging-based psychiatric research, such as the diagnosis of psychiatric diseases, delineation of disease subtypes, normative modeling, and the development of neuroimaging biomarkers, we discuss current challenges. This includes for example the difficulty of training models on small, heterogeneous and biased data sets, the lack of validity of clinical labels, algorithmic bias, and the influence of confounding variables.

Deep learning detection of informative features in tau PET for Alzheimer's disease classification

Background

Alzheimer’s disease (AD) is the most common type of dementia, typically characterized by memory loss followed by progressive cognitive decline and functional impairment. Many clinical trials of potential therapies for AD have failed, and there is currently no approved disease-modifying treatment. Biomarkers for early detection and mechanistic understanding of disease course are critical for drug development and clinical trials. Amyloid has been the focus of most biomarker research. Here, we developed a deep learning-based framework to identify informative features for AD classification using tau positron emission tomography (PET) scans.

Results

The 3D convolutional neural network (CNN)-based classification model of AD from cognitively normal (CN) yielded an average accuracy of 90.8% based on five-fold cross-validation. The LRP model identified the brain regions in tau PET images that contributed most to the AD classification from CN. The top identified regions included the hippocampus, parahippocampus, thalamus, and fusiform. The layer-wise relevance propagation (LRP) results were consistent with those from the voxel-wise analysis in SPM12, showing significant focal AD associated regional tau deposition in the bilateral temporal lobes including the entorhinal cortex. The AD probability scores calculated by the classifier were correlated with brain tau deposition in the medial temporal lobe in MCI participants (r = 0.43 for early MCI and r = 0.49 for late MCI).

Conclusion

A deep learning framework combining 3D CNN and LRP algorithms can be used with tau PET images to identify informative features for AD classification and may have application for early detection during prodromal stages of AD.

Artificial intelligence for brain diseases: A systematic review

Artificial intelligence (AI) is a major branch of computer science that is fruitfully used for analyzing complex medical data and extracting meaningful relationships in datasets, for several clinical aims. Specifically, in the brain care domain, several innovative approaches have achieved remarkable results and open new perspectives in terms of diagnosis, planning, and outcome prediction. In this work, we present an overview of different artificial intelligent techniques used in the brain care domain, along with a review of important clinical applications. A systematic and careful literature search in major databases such as Pubmed, Scopus, and Web of Science was carried out using "artificial intelligence" and "brain" as main keywords. Further references were integrated by cross-referencing from key articles. 155 studies out of 2696 were identified, which actually made use of AI algorithms for different purposes (diagnosis, surgical treatment, intra-operative assistance, and postoperative assessment). Artificial neural networks have risen to prominent positions among the most widely used analytical tools. Classic machine learning approaches such as support vector machine and random forest are still widely used. Task-specific algorithms are designed for solving specific problems. Brain images are one of the most used data types. AI has the possibility to improve clinicians' decision-making ability in neuroscience applications. However, major issues still need to be addressed for a better practical use of AI in the brain. To this aim, it is important to both gather comprehensive data and build explainable AI algorithms.

Application of deep learning in detecting neurological disorders from magnetic resonance images: a survey on the detection of Alzheimer's disease, Parkinson's disease and schizophrenia

Neuroimaging, in particular magnetic resonance imaging (MRI), has been playing an important role in understanding brain functionalities and its disorders during the last couple of decades. These cutting-edge MRI scans, supported by high-performance computational tools and novel ML techniques, have opened up possibilities to unprecedentedly identify neurological disorders. However, similarities in disease phenotypes make it very difficult to detect such disorders accurately from the acquired neuroimaging data. This article critically examines and compares performances of the existing deep learning (DL)-based methods to detect neurological disorders-focusing on Alzheimer's disease, Parkinson's disease and schizophrenia-from MRI data acquired using different modalities including functional and structural MRI. The comparative performance analysis of various DL architectures across different disorders and imaging modalities suggests that the Convolutional Neural Network outperforms other methods in detecting neurological disorders. Towards the end, a number of current research challenges are indicated and some possible future research directions are provided.

Molecular Image-Based Prediction Models of Nuclear Receptor Agonists and Antagonists Using the DeepSnap-Deep Learning Approach with the Tox21 10K Library

The interaction of nuclear receptors (NRs) with chemical compounds can cause dysregulation of endocrine signaling pathways, leading to adverse health outcomes due to the disruption of natural hormones. Thus, identifying possible ligands of NRs is a crucial task for understanding the adverse outcome pathway (AOP) for human toxicity as well as the development of novel drugs. However, the experimental assessment of novel ligands remains expensive and time-consuming. Therefore, an in silico approach with a wide range of applications instead of experimental examination is highly desirable. The recently developed novel molecular image-based deep learning (DL) method, DeepSnap-DL, can produce multiple snapshots from three-dimensional (3D) chemical structures and has achieved high performance in the prediction of chemicals for toxicological evaluation. In this study, we used DeepSnap-DL to construct prediction models of 35 agonist and antagonist allosteric modulators of NRs for chemicals derived from the Tox21 10K library. We demonstrate the high performance of DeepSnap-DL in constructing prediction models. These findings may aid in interpreting the key molecular events of toxicity and support the development of new fields of machine learning to identify environmental chemicals with the potential to interact with NR signaling pathways.

The Latest Trends in Attention Mechanisms and Their Application in Medical Imaging

딥러닝 기술은 빅데이터 및 컴퓨팅 파워를 기반으로 최근 영상의학 분야의 연구에서 괄목할만한 성과를 이루어 내고 있다. 하지만 성능 향상을 위해 딥러닝 네트워크가 깊어질수록 그 내부의 계산 과정을 해석하기 어려워졌는데, 이는 환자의 생명과 직결되는 의료분야의 의사 결정 과정에서는 매우 심각한 문제이다. 이를 해결하기 위해 “설명 가능한 인공지능 기술”이 연구되고 있으며, 그중 하나로 개발된 것이 바로 어텐션(attention) 기법이다. 본 종설에서는 이미 학습이 완료된 네트워크를 분석하기 위한 Post-hoc attention과, 네트워크 성능의 추가적인 향상을 위한 Trainable attention 두 종류의 기법에 대해 각각의 방법 및 의료 영상 연구에 적용된 사례, 그리고 향후 전망 등에 대해 자세히 다루고자 한다.

Uncovering convolutional neural network decisions for diagnosing multiple sclerosis on conventional MRI using layer-wise relevance propagation

Machine learning-based imaging diagnostics has recently reached or even surpassed the level of clinical experts in several clinical domains. However, classification decisions of a trained machine learning system are typically non-transparent, a major hindrance for clinical integration, error tracking or knowledge discovery. In this study, we present a transparent deep learning framework relying on 3D convolutional neural networks (CNNs) and layer-wise relevance propagation (LRP) for diagnosing multiple sclerosis (MS), the most widespread autoimmune neuroinflammatory disease. MS is commonly diagnosed utilizing a combination of clinical presentation and conventional magnetic resonance imaging (MRI), specifically the occurrence and presentation of white matter lesions in T2-weighted images. We hypothesized that using LRP in a naive predictive model would enable us to uncover relevant image features that a trained CNN uses for decision-making. Since imaging markers in MS are well-established this would enable us to validate the respective CNN model. First, we pre-trained a CNN on MRI data from the Alzheimer's Disease Neuroimaging Initiative (n = 921), afterwards specializing the CNN to discriminate between MS patients (n = 76) and healthy controls (n = 71). Using LRP, we then produced a heatmap for each subject in the holdout set depicting the voxel-wise relevance for a particular classification decision. The resulting CNN model resulted in a balanced accuracy of 87.04% and an area under the curve of 96.08% in a receiver operating characteristic curve. The subsequent LRP visualization revealed that the CNN model focuses indeed on individual lesions, but also incorporates additional information such as lesion location, non-lesional white matter or gray matter areas such as the thalamus, which are established conventional and advanced MRI markers in MS. We conclude that LRP and the proposed framework have the capability to make diagnostic decisions of CNN models transparent, which could serve to justify classification decisions for clinical review, verify diagnosis-relevant features and potentially gather new disease knowledge.