Impact of bias field correction on 0.35 T pelvic MR images: evaluation on generative adversarial network-based OARs' auto-segmentation and visual grading assessment.

Purpose: Magnetic resonance imaging (MRI)-guided radiotherapy enables adaptive treatment plans based on daily anatomical changes and accurate organ visualization. However, the bias field artifact can compromise image quality, affecting diagnostic accuracy and quantitative analyses. This study aims to assess the impact of bias field correction on 0.35 T pelvis MRIs by evaluating clinical anatomy visualization and generative adversarial network (GAN) auto-segmentation performance. Materials and methods: 3D simulation MRIs from 60 prostate cancer patients treated on MR-Linac (0.35 T) were collected and preprocessed with the N4ITK algorithm for bias field correction. A 3D GAN architecture was trained, validated, and tested on 40, 10, and 10 patients, respectively, to auto-segment the organs at risk (OARs) rectum and bladder. The GAN was trained and evaluated either with the original or the bias-corrected MRIs. The Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95th) were computed for the segmented volumes of each patient. The Wilcoxon signed-rank test assessed the statistical difference of the metrics within OARs, both with and without bias field correction. Five radiation oncologists blindly scored 22 randomly chosen patients in terms of overall image quality and visibility of boundaries (prostate, rectum, bladder, seminal vesicles) of the original and bias-corrected MRIs. Bennett's S score and Fleiss' kappa were used to assess the pairwise interrater agreement and the interrater agreement among all the observers, respectively. Results: In the test set, the GAN trained and evaluated on original and bias-corrected MRIs showed DSC/HD95th of 0.92/5.63 mm and 0.92/5.91 mm for the bladder and 0.84/10.61 mm and 0.83/9.71 mm for the rectum. No statistical differences in the distribution of the evaluation metrics were found neither for the bladder (DSC: p = 0.07; HD95th: p = 0.35) nor for the rectum (DSC: p = 0.32; HD95th: p = 0.63). From the clinical visual grading assessment, the bias-corrected MRI resulted mostly in either no change or an improvement of the image quality and visualization of the organs' boundaries compared with the original MRI. Conclusion: The bias field correction did not improve the anatomy visualization from a clinical point of view and the OARs' auto-segmentation outputs generated by the GAN.

Comparison of deep learning networks for fully automated head and neck tumor delineation on multi-centric PET/CT images.

OBJECTIVES: Deep learning-based auto-segmentation of head and neck cancer (HNC) tumors is expected to have better reproducibility than manual delineation. Positron emission tomography (PET) and computed tomography (CT) are commonly used in tumor segmentation. However, current methods still face challenges in handling whole-body scans where a manual selection of a bounding box may be required. Moreover, different institutions might still apply different guidelines for tumor delineation. This study aimed at exploring the auto-localization and segmentation of HNC tumors from entire PET/CT scans and investigating the transferability of trained baseline models to external real world cohorts. METHODS: We employed 2D Retina Unet to find HNC tumors from whole-body PET/CT and utilized a regular Unet to segment the union of the tumor and involved lymph nodes. In comparison, 2D/3D Retina Unets were also implemented to localize and segment the same target in an end-to-end manner. The segmentation performance was evaluated via Dice similarity coefficient (DSC) and Hausdorff distance 95th percentile (HD95). Delineated PET/CT scans from the HECKTOR challenge were used to train the baseline models by 5-fold cross-validation. Another 271 delineated PET/CTs from three different institutions (MAASTRO, CRO, BERLIN) were used for external testing. Finally, facility-specific transfer learning was applied to investigate the improvement of segmentation performance against baseline models. RESULTS: Encouraging localization results were observed, achieving a maximum omnidirectional tumor center difference lower than 6.8 cm for external testing. The three baseline models yielded similar averaged cross-validation (CV) results with a DSC in a range of 0.71-0.75, while the averaged CV HD95 was 8.6, 10.7 and 9.8 mm for the regular Unet, 2D and 3D Retina Unets, respectively. More than a 10% drop in DSC and a 40% increase in HD95 were observed if the baseline models were tested on the three external cohorts directly. After the facility-specific training, an improvement in external testing was observed for all models. The regular Unet had the best DSC (0.70) for the MAASTRO cohort, and the best HD95 (7.8 and 7.9 mm) in the MAASTRO and CRO cohorts. The 2D Retina Unet had the best DSC (0.76 and 0.67) for the CRO and BERLIN cohorts, and the best HD95 (12.4 mm) for the BERLIN cohort. CONCLUSION: The regular Unet outperformed the other two baseline models in CV and most external testing cohorts. Facility-specific transfer learning can potentially improve HNC segmentation performance for individual institutions, where the 2D Retina Unets could achieve comparable or even better results than the regular Unet.

Real-time motion management in MRI-guided radiotherapy: Current status and AI-enabled prospects.

MRI-guided radiotherapy (MRIgRT) is a highly complex treatment modality, allowing adaptation to anatomical changes occurring from one treatment day to the other (inter-fractional), but also to motion occurring during a treatment fraction (intra-fractional). In this vision paper, we describe the different steps of intra-fractional motion management during MRIgRT, from imaging to beam adaptation, and the solutions currently available both clinically and at a research level. Furthermore, considering the latest developments in the literature, a workflow is foreseen in which motion-induced over- and/or under-dosage is compensated in 3D, with minimal impact to the radiotherapy treatment time. Considering the time constraints of real-time adaptation, a particular focus is put on artificial intelligence (AI) solutions as a fast and accurate alternative to conventional algorithms.

Experimental comparison of linear regression and LSTM motion prediction models for MLC-tracking on an MRI-linac.

BACKGROUND: Magnetic resonance imaging (MRI)-guided radiotherapy with multileaf collimator (MLC)-tracking is a promising technique for intra-fractional motion management, achieving high dose conformality without prolonging treatment times. To improve beam-target alignment, the geometric error due to system latency should be reduced by using temporal prediction. PURPOSE: To experimentally compare linear regression (LR) and long-short-term memory (LSTM) motion prediction models for MLC-tracking on an MRI-linac using multiple patient-derived traces with different complexities. METHODS: Experiments were performed on a prototype 1.0 T MRI-linac capable of MLC-tracking. A motion phantom was programmed to move a target in superior-inferior (SI) direction according to eight lung cancer patient respiratory motion traces. Target centroid positions were localized from sagittal 2D cine MRIs acquired at 4 Hz using a template matching algorithm. The centroid positions were input to one of four motion prediction models. We used (1) a LSTM network which had been optimized in a previous study on patient data from another cohort (offline LSTM). We also used (2) the same LSTM model as a starting point for continuous re-optimization of its weights during the experiment based on recent motion (offline+online LSTM). Furthermore, we implemented (3) a continuously updated LR model, which was solely based on recent motion (online LR). Finally, we used (4) the last available target centroid without any changes as a baseline (no-predictor). The predictions of the models were used to shift the MLC aperture in real-time. An electronic portal imaging device (EPID) was used to visualize the target and MLC aperture during the experiments. Based on the EPID frames, the root-mean-square error (RMSE) between the target and the MLC aperture positions was used to assess the performance of the different motion predictors. Each combination of motion trace and prediction model was repeated twice to test stability, for a total of 64 experiments. RESULTS: The end-to-end latency of the system was measured to be (389 ± 15) ms and was successfully mitigated by both LR and LSTM models. The offline+online LSTM was found to outperform the other models for all investigated motion traces. It obtained a median RMSE over all traces of (2.8 ± 1.3) mm, compared to the (3.2 ± 1.9) mm of the offline LSTM, the (3.3 ± 1.4) mm of the online LR and the (4.4 ± 2.4) mm when using the no-predictor. According to statistical tests, differences were significant (p-value <0.05) among all models in a pair-wise comparison, but for the offline LSTM and online LR pair. The offline+online LSTM was found to be more reproducible than the offline LSTM and the online LR with a maximum deviation in RMSE between two measurements of 10%. CONCLUSIONS: This study represents the first experimental comparison of different prediction models for MRI-guided MLC-tracking using several patient-derived respiratory motion traces. We have shown that among the investigated models, continuously re-optimized LSTM networks are the most promising to account for the end-to-end system latency in MRI-guided radiotherapy with MLC-tracking.

ScatterNet for projection-based 4D cone-beam computed tomography intensity correction of lung cancer patients.

Background and purpose: In radiotherapy, dose calculations based on 4D cone beam CTs (4DCBCTs) require image intensity corrections. This retrospective study compared the dose calculation accuracy of a deep learning, projection-based scatter correction workflow (ScatterNet), to slower workflows: conventional 4D projection-based scatter correction (CBCTcor) and a deformable image registration (DIR)-based method (4DvCT). Materials and methods: For 26 lung cancer patients, planning CTs (pCTs), 4DCTs and CBCT projections were available. ScatterNet was trained with pairs of raw and corrected CBCT projections. Corrected projections from ScatterNet and the conventional workflow were reconstructed using MA-ROOSTER, yielding 4DCBCTSN and 4DCBCTcor. The 4DvCT was generated by 4DCT to 4DCBCT DIR, as part of the 4DCBCTcor workflow. Robust intensity modulated proton therapy treatment plans were created on free-breathing pCTs. 4DCBCTSN was compared to 4DCBCTcor and the 4DvCT in terms of image quality and dose calculation accuracy (dose-volume-histogram parameters and 3%/3mm gamma analysis). Results: 4DCBCTSN resulted in an average mean absolute error of 87HU and 102HU when compared to 4DCBCTcor and 4DvCT respectively. High agreement was observed in targets with median dose differences of 0.4Gy (4DCBCTSN-4DCBCTcor) and 0.3Gy (4DCBCTSN-4DvCT). The gamma analysis showed high average 3%/3mm pass rates of 96% for both 4DCBCTSN vs. 4DCBCTcor and 4DCBCTSN vs. 4DvCT. Conclusions: Accurate 4D dose calculations are feasible for lung cancer patients using ScatterNet for 4DCBCT correction. Average scatter correction times could be reduced from 10min (4DCBCTcor) to 3.9s, showing the clinical suitability of the proposed deep learning-based method.

Planning CT Identifies Patients at Risk of High Prostate Intrafraction Motion.

Prostate motion (standard deviation, range of motion, and diffusion coefficient) was calculated from 4D ultrasound data of 1791 fractions of radiation therapy in N = 100 patients. The inner diameter of the lesser pelvis was obtained from transversal slices through the pubic symphysis in planning CTs. On the lateral and craniocaudal axes, motility increases significantly (t-test, p < 0.005) with the inner diameter of the lesser pelvis. A diameter of >106 mm (ca. 6th decile) is a good predictor for high prostate intrafraction motion (ca. 9th decile). The corresponding area under the receiver operator curve (AUROC) is 80% in the lateral direction, 68% to 80% in the craniocaudal direction, and 62% to 70% in the vertical direction. On the lateral x-axis, the proposed test is 100% sensitive and has a 100% negative predictive value for all three characteristics (standard deviation, range of motion, and diffusion coefficient). On the craniocaudal z-axis, the proposed test is 79% to 100% sensitive and reaches 95% to 100% negative predictive value. On the vertical axis, the proposed test still delivers 98% negative predictive value but is not particularly sensitive. Overall, the proposed predictor is able to help identify patients at risk of high prostate motion based on a single planning CT.

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.

Evaluation of real-time tumor contour prediction using LSTM networks for MR-guided radiotherapy.

BACKGROUND AND PURPOSE: Magnetic resonance imaging guided radiotherapy (MRgRT) with deformable multileaf collimator (MLC) tracking would allow to tackle both rigid displacement and tumor deformation without prolonging treatment. However, the system latency must be accounted for by predicting future tumor contours in real-time. We compared the performance of three artificial intelligence (AI) algorithms based on long short-term memory (LSTM) modules for the prediction of 2D-contours 500ms into the future. MATERIALS AND METHODS: Models were trained (52 patients, 3.1h of motion), validated (18 patients, 0.6h) and tested (18 patients, 1.1h) with cine MRs from patients treated at one institution. Additionally, we used three patients (2.9h) treated at another institution as second testing set. We implemented 1) a classical LSTM network (LSTM-shift) predicting tumor centroid positions in superior-inferior and anterior-posterior direction which are used to shift the last observed tumor contour. The LSTM-shift model was optimized both in an offline and online fashion. We also implemented 2) a convolutional LSTM model (ConvLSTM) to directly predict future tumor contours and 3) a convolutional LSTM combined with spatial transformer layers (ConvLSTM-STL) to predict displacement fields used to warp the last tumor contour. RESULTS: The online LSTM-shift model was found to perform slightly better than the offline LSTM-shift and significantly better than the ConvLSTM and ConvLSTM-STL. It achieved a 50% Hausdorff distance of 1.2mm and 1.0mm for the two testing sets, respectively. Larger motion ranges were found to lead to more substantial performance differences across the models. CONCLUSION: LSTM networks predicting future centroids and shifting the last tumor contour are the most suitable for tumor contour prediction. The obtained accuracy would allow to reduce residual tracking errors during MRgRT with deformable MLC-tracking.

DeepClassPathway: Molecular pathway aware classification using explainable deep learning.

OBJECTIVE: HPV-associated head and neck cancer is correlated with favorable prognosis; however, its underlying biology is not fully understood. We propose an explainable convolutional neural network (CNN) classifier, DeepClassPathway, that predicts HPV-status and allows patient-specific identification of molecular pathways driving classifier decisions. METHODS: The CNN was trained to classify HPV-status on transcriptome data from 264 (13% HPV-positive) and tested on 85 (25% HPV-positive) head and neck squamous carcinoma patients after transformation into 2D-treemaps representing molecular pathways. Grad-CAM saliency was used to quantify pathways contribution to individual CNN decisions. Model stability was assessed by shuffling pathways within 2D-images. RESULTS: The classification performance of the CNN-ensembles achieved ROC-AUC/PR-AUC of 0.96/0.90 for all treemap variants. Quantification of the averaged pathway saliency heatmaps consistently identified KRAS, spermatogenesis, bile acid metabolism, and inflammation signaling pathways as the four most informative for classifying HPV-positive patients and MYC targets, epithelial-mesenchymal transition, and protein secretion pathways for HPV-negative patients. CONCLUSION: We have developed and applied an explainable CNN classification approach to transcriptome data from an oncology cohort with typical sample size that allows classification while accounting for the importance of molecular pathways in individual-level decisions.

18F-Fluorodeoxyglucose Positron Emission Tomography of Head and Neck Cancer: Location and HPV Specific Parameters for Potential Treatment Individualization.

Purpose: 18F-fluorodeoxyglucose positron emission tomography (FDG-PET) is utilized for staging and treatment planning of head and neck squamous cell carcinomas (HNSCC). Some older publications on the prognostic relevance showed inconclusive results, most probably due to small study sizes. This study evaluates the prognostic and potentially predictive value of FDG-PET in a large multi-center analysis. Methods: Original analysis of individual FDG-PET and patient data from 16 international centers (8 institutional datasets, 8 public repositories) with 1104 patients. All patients received curative intent radiotherapy/chemoradiation (CRT) and pre-treatment FDG-PET imaging. Primary tumors were semi-automatically delineated for calculation of SUVmax, SUVmean, metabolic tumor volume (MTV) and total lesion glycolysis (TLG). Cox regression analyses were performed for event-free survival (EFS), overall survival (OS), loco-regional control (LRC) and freedom from distant metastases (FFDM). Results: FDG-PET parameters were associated with patient outcome in the whole cohort regarding clinical endpoints (EFS, OS, LRC, FFDM), in uni- and multivariate Cox regression analyses. Several previously published cut-off values were successfully validated. Subgroup analyses identified tumor- and human papillomavirus (HPV) specific parameters. In HPV positive oropharynx cancer (OPC) SUVmax was well suited to identify patients with excellent LRC for organ preservation. Patients with SUVmax of 14 or less were unlikely to develop loco-regional recurrence after definitive CRT. In contrast FDG PET parameters deliver only limited prognostic information in laryngeal cancer. Conclusion: FDG-PET parameters bear considerable prognostic value in HNSCC and potential predictive value in subgroups of patients, especially regarding treatment de-intensification and organ-preservation. The potential predictive value needs further validation in appropriate control groups. Further research on advanced imaging approaches including radiomics or artificial intelligence methods should implement the identified cut-off values as benchmark routine imaging parameters.

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

OBJECTIVES: Recent studies have shown that deep learning based on pre-treatment positron emission tomography (PET) or computed tomography (CT) is promising for distant metastasis (DM) and overall survival (OS) prognosis in head and neck cancer (HNC). However, lesion segmentation is typically required, resulting in a predictive power susceptible to variations in primary and lymph node gross tumor volume (GTV) segmentation. This study aimed at achieving prognosis without GTV segmentation, and extending single modality prognosis to joint PET/CT to allow investigating the predictive performance of combined- compared to single-modality inputs. METHODS: We employed a 3D-Resnet combined with a time-to-event outcome model to incorporate censoring information. We focused on the prognosis of DM and OS for HNC patients. For each clinical endpoint, five models with PET and/or CT images as input were compared: PET-GTV, PET-only, CT-GTV, CT-only, and PET/CT-GTV models, where -GTV indicates that the corresponding images were masked using the GTV contour. Publicly available delineated CT and PET scans from 4 different Canadian hospitals (293) and the MAASTRO clinic (74) were used for training by 3-fold cross-validation (CV). For independent testing, we used 110 patients from a collaborating institution. The predictive performance was evaluated via Harrell's Concordance Index (HCI) and Kaplan-Meier curves. RESULTS: In a 5-year time-to-event analysis, all models could produce CV HCIs with median values around 0.8 for DM and 0.7 for OS. The best performance was obtained with the PET-only model, achieving a median testing HCI of 0.82 for DM and 0.69 for OS. Compared with the PET/CT-GTV model, the PET-only still had advantages of up to 0.07 in terms of testing HCI. The Kaplan-Meier curves and corresponding log-rank test results also demonstrated significant stratification capability of our models for the testing cohort. CONCLUSION: Deep learning-based DM and OS time-to-event models showed predictive capability and could provide indications for personalized RT. The best predictive performance achieved by the PET-only model suggested GTV segmentation might be less relevant for PET-based prognosis.

Offline and online LSTM networks for respiratory motion prediction in MR-guided radiotherapy.

Objective.Gated beam delivery is the current clinical practice for respiratory motion compensation in MR-guided radiotherapy, and further research is ongoing to implement tracking. To manage intra-fractional motion using multileaf collimator tracking the total system latency needs to be accounted for in real-time. In this study, long short-term memory (LSTM) networks were optimized for the prediction of superior-inferior tumor centroid positions extracted from clinically acquired 2D cine MRIs.Approach.We used 88 patients treated at the University Hospital of the LMU Munich for training and validation (70 patients, 13.1 h), and for testing (18 patients, 3.0 h). Three patients treated at Fondazione Policlinico Universitario Agostino Gemelli were used as a second testing set (1.5 h). The performance of the LSTMs in terms of root mean square error (RMSE) was compared to baseline linear regression (LR) models for forecasted time spans of 250 ms, 500 ms and 750 ms. Both the LSTM and the LR were trained with offline (offlineLSTM andofflineLR) and online schemes (offline+onlineLSTM andonlineLR), the latter to allow for continuous adaptation to recent respiratory patterns.Main results.We found theoffline+onlineLSTM to perform best for all investigated forecasts. Specifically, when predicting 500 ms ahead it achieved a mean RMSE of 1.20 mm and 1.00 mm, while the best performing LR model achieved a mean RMSE of 1.42 mm and 1.22 mm for the LMU and Gemelli testing set, respectively.Significance.This indicates that LSTM networks have potential as respiratory motion predictors and that continuous online re-optimization can enhance their performance.

MR-guided radiotherapy in node-positive non-small cell lung cancer and severely limited pulmonary reserve: a report proposing a new clinical pathway for the management of high-risk patients.

INTRODUCTION: Online MR-guided radiotherapy (MRgRT) is a relatively novel advancement in the field of radiation oncology, ensuring superior soft-tissue visualisation, allowing for online plan adaptation to anatomical and functional interfractional changes and improved motion management. Platinum-based chemoradiation followed by durvalumab is the recommended treatment for stage IIB(N1)/III NSCLC. However, this is only the case for patients with favourable risk factors and sufficient pulmonary function and reserve. METHODS: Herein, we present a technical report on tumour motion and breathing curve analyses of the first patient with node-positive stage IIB NSCLC and severely compromised pulmonary function and reserve [total lung capacity (TLC) 8.78L/132% predicted, residual volume (RV) 6.35L/271% predicted, vital capacity (VC) max 2.43L/58% predicted, FEV1 1.19L/38% predicted, DLCO-SB corrected for hemoglobin 2.76 mmol/min/kPa/30% predicted] treated in a prospective observational study with moderately hypofractionated MRgRT to a total dose of 48.0 Gy/16 daily fractions on the MRIdian system (Viewray Inc, Oakwood, USA). RESULTS: Radiotherapy was well tolerated with no relevant toxicity. First follow-up imaging at 3 months post-radiotherapy showed a partial remission. The distinctive features of this case are the patient's severely compromised pulmonary function and the first online MR-guided accelerated hypofractionated radiotherapy treatment for primary node-positive NSCLC. CONCLUSIONS: This technical report describes the first patient treated in a prospective observational study evaluating the feasibility of this relatively novel technology in stage IIB(N1)/III disease, proposing a clinical pathway for the management of high-risk patients.

Risk Stratification Using 18F-FDG PET/CT and Artificial Neural Networks in Head and Neck Cancer Patients Undergoing Radiotherapy.

This study retrospectively analyzed the performance of artificial neural networks (ANN) to predict overall survival (OS) or locoregional failure (LRF) in HNSCC patients undergoing radiotherapy, based on 2-[18F]FDG PET/CT and clinical covariates. We compared predictions relying on three different sets of features, extracted from 230 patients. Specifically, (i) an automated feature selection method independent of expert rating was compared with (ii) clinical variables with proven influence on OS or LRF and (iii) clinical data plus expert-selected SUV metrics. The three sets were given as input to an artificial neural network for outcome prediction, evaluated by Harrell's concordance index (HCI) and by testing stratification capability. For OS and LRF, the best performance was achieved with expert-based PET-features (0.71 HCI) and clinical variables (0.70 HCI), respectively. For OS stratification, all three feature sets were significant, whereas for LRF only expert-based PET-features successfully classified low vs. high-risk patients. Based on 2-[18F]FDG PET/CT features, stratification into risk groups using ANN for OS and LRF is possible. Differences in the results for different feature sets confirm the relevance of feature selection, and the key importance of expert knowledge vs. automated selection.

Distant metastasis time to event analysis with CNNs in independent head and neck cancer cohorts.

Deep learning models based on medical images play an increasingly important role for cancer outcome prediction. The standard approach involves usage of convolutional neural networks (CNNs) to automatically extract relevant features from the patient's image and perform a binary classification of the occurrence of a given clinical endpoint. In this work, a 2D-CNN and a 3D-CNN for the binary classification of distant metastasis (DM) occurrence in head and neck cancer patients were extended to perform time-to-event analysis. The newly built CNNs incorporate censoring information and output DM-free probability curves as a function of time for every patient. In total, 1037 patients were used to build and assess the performance of the time-to-event model. Training and validation was based on 294 patients also used in a previous benchmark classification study while for testing 743 patients from three independent cohorts were used. The best network could reproduce the good results from 3-fold cross validation [Harrell's concordance indices (HCIs) of 0.78, 0.74 and 0.80] in two out of three testing cohorts (HCIs of 0.88, 0.67 and 0.77). Additionally, the capability of the models for patient stratification into high and low-risk groups was investigated, the CNNs being able to significantly stratify all three testing cohorts. Results suggest that image-based deep learning models show good reliability for DM time-to-event analysis and could be used for treatment personalisation.