Deep learning based time-to-event analysis with PET, CT and joint PET/CT for head and neck cancer prognosis

Applications and challenges of neural networks in otolaryngology (Review)

Artificial Intelligence (AI) has become a topic of interest that is frequently debated in all research fields. The medical field is no exception, where several unanswered questions remain. When and how this field can benefit from AI support in daily routines are the most frequently asked questions. The present review aims to present the types of neural networks (NNs) available for development, discussing their advantages, disadvantages and how they can be applied practically. In addition, the present review summarizes how NNs (combined with various other features) have already been applied in studies in the ear nose throat research field, from assisting diagnosis to treatment management. Although the answer to this question regarding AI remains elusive, understanding the basics and types of applicable NNs can lead to future studies possibly using more than one type of NN. This approach may bypass the actual limitations in accuracy and relevance of information generated by AI. The proposed studies, the majority of which used convolutional NNs, obtained accuracies varying 70-98%, with a number of studies having the AI trained on a limited number of cases (<100 patients). The lack of standardization in AI protocols for research negatively affects data homogeneity and transparency of databases.

BAF-Net: bidirectional attention-aware fluid pyramid feature integrated multimodal fusion network for diagnosis and prognosis

Objective. To go beyond the deficiencies of the three conventional multimodal fusion strategies (i.e. input-, feature- and output-level fusion), we propose a bidirectional attention-aware fluid pyramid feature integrated fusion network (BAF-Net) with cross-modal interactions for multimodal medical image diagnosis and prognosis.Approach. BAF-Net is composed of two identical branches to preserve the unimodal features and one bidirectional attention-aware distillation stream to progressively assimilate cross-modal complements and to learn supplementary features in both bottom-up and top-down processes. Fluid pyramid connections were adopted to integrate the hierarchical features at different levels of the network, and channel-wise attention modules were exploited to mitigate cross-modal cross-level incompatibility. Furthermore, depth-wise separable convolution was introduced to fuse the cross-modal cross-level features to alleviate the increase in parameters to a great extent. The generalization abilities of BAF-Net were evaluated in terms of two clinical tasks: (1) an in-house PET-CT dataset with 174 patients for differentiation between lung cancer and pulmonary tuberculosis. (2) A public multicenter PET-CT head and neck cancer dataset with 800 patients from nine centers for overall survival prediction.Main results. On the LC-PTB dataset, improved performance was found in BAF-Net (AUC = 0.7342) compared with input-level fusion model (AUC = 0.6825;p< 0.05), feature-level fusion model (AUC = 0.6968;p= 0.0547), output-level fusion model (AUC = 0.7011;p< 0.05). On the H&N cancer dataset, BAF-Net (C-index = 0.7241) outperformed the input-, feature-, and output-level fusion model, with 2.95%, 3.77%, and 1.52% increments of C-index (p= 0.3336, 0.0479 and 0.2911, respectively). The ablation experiments demonstrated the effectiveness of all the designed modules regarding all the evaluated metrics in both datasets.Significance. Extensive experiments on two datasets demonstrated better performance and robustness of BAF-Net than three conventional fusion strategies and PET or CT unimodal network in terms of diagnosis and prognosis.

Deep learning-based outcome prediction using PET/CT and automatically predicted probability maps of primary tumor in patients with oropharyngeal cancer

BACKGROUND AND OBJECTIVE

Recently, deep learning (DL) algorithms showed to be promising in predicting outcomes such as distant metastasis-free survival (DMFS) and overall survival (OS) using pre-treatment imaging in head and neck cancer. Gross Tumor Volume of the primary tumor (GTVp) segmentation is used as an additional channel in the input to DL algorithms to improve model performance. However, the binary segmentation mask of the GTVp directs the focus of the network to the defined tumor region only and uniformly. DL models trained for tumor segmentation have also been used to generate predicted tumor probability maps (TPM) where each pixel value corresponds to the degree of certainty of that pixel to be classified as tumor. The aim of this study was to explore the effect of using TPM as an extra input channel of CT- and PET-based DL prediction models for oropharyngeal cancer (OPC) patients in terms of local control (LC), regional control (RC), DMFS and OS.

METHODS

We included 399 OPC patients from our institute that were treated with definitive (chemo)radiation. For each patient, CT and PET scans and GTVp contours, used for radiotherapy treatment planning, were collected. We first trained a previously developed 2.5D DL framework for tumor probability prediction by 5-fold cross validation using 131 patients. Then, a 3D ResNet18 was trained for outcome prediction using the 3D TPM as one of the possible inputs. The endpoints were LC, RC, DMFS, and OS. We performed 3-fold cross validation on 168 patients for each endpoint using different combinations of image modalities as input. The final prediction in the test set (100) was obtained by averaging the predictions of the 3-fold models. The C-index was used to evaluate the discriminative performance of the models.

RESULTS

The models trained replacing the GTVp contours with the TPM achieved the highest C-indexes for LC (0.74) and RC (0.60) prediction. For OS, using the TPM or the GTVp as additional image modality resulted in comparable C-indexes (0.72 and 0.74).

CONCLUSIONS

Adding predicted TPMs instead of GTVp contours as an additional input channel for DL-based outcome prediction models improved model performance for LC and RC.

Deep Learning Techniques and Imaging in Otorhinolaryngology-A State-of-the-Art Review

Over the last decades, the field of medicine has witnessed significant progress in artificial intelligence (AI), the Internet of Medical Things (IoMT), and deep learning (DL) systems. Otorhinolaryngology, and imaging in its various subspecialties, has not remained untouched by this transformative trend. As the medical landscape evolves, the integration of these technologies becomes imperative in augmenting patient care, fostering innovation, and actively participating in the ever-evolving synergy between computer vision techniques in otorhinolaryngology and AI. To that end, we conducted a thorough search on MEDLINE for papers published until June 2023, utilizing the keywords 'otorhinolaryngology', 'imaging', 'computer vision', 'artificial intelligence', and 'deep learning', and at the same time conducted manual searching in the references section of the articles included in our manuscript. Our search culminated in the retrieval of 121 related articles, which were subsequently subdivided into the following categories: imaging in head and neck, otology, and rhinology. Our objective is to provide a comprehensive introduction to this burgeoning field, tailored for both experienced specialists and aspiring residents in the domain of deep learning algorithms in imaging techniques in otorhinolaryngology.

Comparison of computed tomography image features extracted by radiomics, self-supervised learning and end-to-end deep learning for outcome prediction of oropharyngeal cancer

Background and purpose

To compare the prediction performance of image features of computed tomography (CT) images extracted by radiomics, self-supervised learning and end-to-end deep learning for local control (LC), regional control (RC), locoregional control (LRC), distant metastasis-free survival (DMFS), tumor-specific survival (TSS), overall survival (OS) and disease-free survival (DFS) of oropharyngeal squamous cell carcinoma (OPSCC) patients after (chemo)radiotherapy.

Methods and materials

The OPC-Radiomics dataset was used for model development and independent internal testing and the UMCG-OPC set for external testing. Image features were extracted from the Gross Tumor Volume contours of the primary tumor (GTVt) regions in CT scans when using radiomics or a self-supervised learning-based method (autoencoder). Clinical and combined (radiomics, autoencoder or end-to-end) models were built using multivariable Cox proportional-hazard analysis with clinical features only and both clinical and image features for LC, RC, LRC, DMFS, TSS, OS and DFS prediction, respectively.

Results

In the internal test set, combined autoencoder models performed better than clinical models and combined radiomics models for LC, RC, LRC, DMFS, TSS and DFS prediction (largest improvements in C-index: 0.91 vs. 0.76 in RC and 0.74 vs. 0.60 in DMFS). In the external test set, combined radiomics models performed better than clinical and combined autoencoder models for all endpoints (largest improvements in LC, 0.82 vs. 0.71). Furthermore, combined models performed better in risk stratification than clinical models and showed good calibration for most endpoints.

Conclusions

Image features extracted using self-supervised learning showed best internal prediction performance while radiomics features have better external generalizability.

Head and neck cancer treatment outcome prediction: a comparison between machine learning with conventional radiomics features and deep learning radiomics

Background

Radiomics can provide in-depth characterization of cancers for treatment outcome prediction. Conventional radiomics rely on extraction of image features within a pre-defined image region of interest (ROI) which are typically fed to a classification algorithm for prediction of a clinical endpoint. Deep learning radiomics allows for a simpler workflow where images can be used directly as input to a convolutional neural network (CNN) with or without a pre-defined ROI.

Purpose

The purpose of this study was to evaluate (i) conventional radiomics and (ii) deep learning radiomics for predicting overall survival (OS) and disease-free survival (DFS) for patients with head and neck squamous cell carcinoma (HNSCC) using pre-treatment 18F-fluorodeoxuglucose positron emission tomography (FDG PET) and computed tomography (CT) images.

Materials and methods

FDG PET/CT images and clinical data of patients with HNSCC treated with radio(chemo)therapy at Oslo University Hospital (OUS; n = 139) and Maastricht University Medical Center (MAASTRO; n = 99) were collected retrospectively. OUS data was used for model training and initial evaluation. MAASTRO data was used for external testing to assess cross-institutional generalizability. Models trained on clinical and/or conventional radiomics features, with or without feature selection, were compared to CNNs trained on PET/CT images without or with the gross tumor volume (GTV) included. Model performance was measured using accuracy, area under the receiver operating characteristic curve (AUC), Matthew's correlation coefficient (MCC), and the F1 score calculated for both classes separately.

Results

CNNs trained directly on images achieved the highest performance on external data for both endpoints. Adding both clinical and radiomics features to these image-based models increased performance further. Conventional radiomics including clinical data could achieve competitive performance. However, feature selection on clinical and radiomics data lead to overfitting and poor cross-institutional generalizability. CNNs without tumor and node contours achieved close to on-par performance with CNNs including contours.

Conclusion

High performance and cross-institutional generalizability can be achieved by combining clinical data, radiomics features and medical images together with deep learning models. However, deep learning models trained on images without contours can achieve competitive performance and could see potential use as an initial screening tool for high-risk patients.

A convolutional neural network with self-attention for fully automated metabolic tumor volume delineation of head and neck cancer in [Formula: see text]F]FDG PET/CT

PURPOSE

PET-derived metabolic tumor volume (MTV) and total lesion glycolysis of the primary tumor are known to be prognostic of clinical outcome in head and neck cancer (HNC). Including evaluation of lymph node metastases can further increase the prognostic value of PET but accurate manual delineation and classification of all lesions is time-consuming and prone to interobserver variability. Our goal, therefore, was development and evaluation of an automated tool for MTV delineation/classification of primary tumor and lymph node metastases in PET/CT investigations of HNC patients.

METHODS

Automated lesion delineation was performed with a residual 3D U-Net convolutional neural network (CNN) incorporating a multi-head self-attention block. 698 [Formula: see text]F]FDG PET/CT scans from 3 different sites and 5 public databases were used for network training and testing. An external dataset of 181 [Formula: see text]F]FDG PET/CT scans from 2 additional sites was employed to assess the generalizability of the network. In these data, primary tumor and lymph node (LN) metastases were interactively delineated and labeled by two experienced physicians. Performance of the trained network models was assessed by 5-fold cross-validation in the main dataset and by pooling results from the 5 developed models in the external dataset. The Dice similarity coefficient (DSC) for individual delineation tasks and the primary tumor/metastasis classification accuracy were used as evaluation metrics. Additionally, a survival analysis using univariate Cox regression was performed comparing achieved group separation for manual and automated delineation, respectively.

RESULTS

In the cross-validation experiment, delineation of all malignant lesions with the trained U-Net models achieves DSC of 0.885, 0.805, and 0.870 for primary tumor, LN metastases, and the union of both, respectively. In external testing, the DSC reaches 0.850, 0.724, and 0.823 for primary tumor, LN metastases, and the union of both, respectively. The voxel classification accuracy was 98.0% and 97.9% in cross-validation and external data, respectively. Univariate Cox analysis in the cross-validation and the external testing reveals that manually and automatically derived total MTVs are both highly prognostic with respect to overall survival, yielding essentially identical hazard ratios (HR) ([Formula: see text]; [Formula: see text] vs. [Formula: see text]; [Formula: see text] in cross-validation and [Formula: see text]; [Formula: see text] vs. [Formula: see text]; [Formula: see text] in external testing).

CONCLUSION

To the best of our knowledge, this work presents the first CNN model for successful MTV delineation and lesion classification in HNC. In the vast majority of patients, the network performs satisfactory delineation and classification of primary tumor and lymph node metastases and only rarely requires more than minimal manual correction. It is thus able to massively facilitate study data evaluation in large patient groups and also does have clear potential for supervised clinical application.

Functional-structural sub-region graph convolutional network (FSGCN): Application to the prognosis of head and neck cancer with PET/CT imaging

BACKGROUND AND OBJECTIVE

Accurate risk stratification is crucial for enabling personalized treatment for head and neck cancer (HNC). Current PET/CT image-based prognostic methods include radiomics analysis and convolutional neural network (CNN), while extracting radiomics or deep features in grid Euclidean space has inherent limitations for risk stratification. Here, we propose a functional-structural sub-region graph convolutional network (FSGCN) for accurate risk stratification of HNC.

METHODS

This study collected 642 patients from 8 different centers in The Cancer Imaging Archive (TCIA), 507 patients from 5 centers were used for training, and 135 patients from 3 centers were used for testing. The tumor was first clustered into multiple sub-regions by using PET and CT voxel information, and radiomics features were extracted from each sub-region to characterize its functional and structural information, a graph was then constructed to format the relationship/difference among different sub-regions in non-Euclidean space for each patient, followed by a residual gated graph convolutional network, the prognostic score was finally generated to predict the progression-free survival (PFS).

RESULTS

In the testing cohort, compared with radiomics or FSGCN or clinical model alone, the model PETCT^Fea_CT^ROI + Cli that integrates FSGCN prognostic score and clinical parameter achieved the highest C-index and AUC of 0.767 (95% CI: 0.759-0.774) and 0.781 (95% CI: 0.774-0.788), respectively for PFS prediction. Besides, it also showed good prognostic performance on the secondary endpoints OS, RFS, and MFS in the testing cohort, with C-index of 0.786 (95% CI: 0.778-0.795), 0.775 (95% CI: 0.767-0.782) and 0.781 (95% CI: 0.772-0.789), respectively.

CONCLUSIONS

The proposed FSGCN can better capture the metabolic or anatomic difference/interaction among sub-regions of the whole tumor imaged with PET/CT. Extensive multi-center experiments demonstrated its capability and generalization of prognosis prediction in HNC over conventional radiomics analysis.

Longitudinal and Multimodal Radiomics Models for Head and Neck Cancer Outcome Prediction

Radiomics analysis provides a promising avenue towards the enabling of personalized radiotherapy. Most frequently, prognostic radiomics models are based on features extracted from medical images that are acquired before treatment. Here, we investigate whether combining data from multiple timepoints during treatment and from multiple imaging modalities can improve the predictive ability of radiomics models. We extracted radiomics features from computed tomography (CT) images acquired before treatment as well as two and three weeks after the start of radiochemotherapy for 55 patients with locally advanced head and neck squamous cell carcinoma (HNSCC). Additionally, we obtained features from FDG-PET images taken before treatment and three weeks after the start of therapy. Cox proportional hazards models were then built based on features of the different image modalities, treatment timepoints, and combinations thereof using two different feature selection methods in a five-fold cross-validation approach. Based on the cross-validation results, feature signatures were derived and their performance was independently validated. Discrimination regarding loco-regional control was assessed by the concordance index (C-index) and log-rank tests were performed to assess risk stratification. The best prognostic performance was obtained for timepoints during treatment for all modalities. Overall, CT was the best discriminating modality with an independent validation C-index of 0.78 for week two and weeks two and three combined. However, none of these models achieved statistically significant patient stratification. Models based on FDG-PET features from week three provided both satisfactory discrimination (C-index = 0.61 and 0.64) and statistically significant stratification (p=0.044 and p<0.001), but produced highly imbalanced risk groups. After independent validation on larger datasets, the value of (multimodal) radiomics models combining several imaging timepoints should be prospectively assessed for personalized treatment strategies.