Sebastian Marschner

Objective. Fast and accurate organ-at-risk (OAR) and gross tumor volume (GTV) contour propagation methods are needed to improve the efficiency of magnetic resonance (MR) imaging-guided radiotherapy. We trained deformable image registration networks to accurately propagate contours from planning to fraction MR images. Approach. Data from 140 stage 1–2 lung cancer patients treated at a 0.35 T MR-Linac were split into 102/17/21 for training/validation/testing. Additionally, 18 central lung tumor patients, treated at a 0.35 T MR-Linac externally, and 14 stage 3 lung cancer patients from a phase 1 clinical trial, treated at 0.35 T or 1.5 T MR-Linacs at three institutions, were used for external testing. Planning and fraction images were paired (490 pairs) for training. Two hybrid transformer-convolutional neural network TransMorph models with mean squared error (MSE), Dice similarity coefficient (DSC), and regularization losses (TM_MSE+Dice) or MSE and regularization losses (TM_MSE) were trained to deformably register planning to fraction images. The TransMorph models predicted diffeomorphic dense displacement fields. Multi-label images including seven thoracic OARs and the GTV were propagated to generate fraction segmentations. Model predictions were compared with contours obtained through B-spline, vendor registration and the auto-segmentation method nnUNet. Evaluation metrics included the DSC and Hausdorff distance percentiles (50th and 95th) against clinical contours. Main results. TM_MSE+Dice and TM_MSE achieved mean OARs/GTV DSCs of 0.90/0.82 and 0.90/0.79 for the internal and 0.84/0.77 and 0.85/0.76 for the central lung tumor external test data. On stage 3 data, TM_MSE+Dice achieved mean OARs/GTV DSCs of 0.87/0.79 and 0.83/0.78 for the 0.35 T MR-Linac datasets, and 0.87/0.75 for the 1.5 T MR-Linac dataset. TM_MSE+Dice and TM_MSE had significantly higher geometric accuracy than other methods on external data. No significant difference between TM_MSE+Dice and TM_MSE was found. Significance. TransMorph models achieved time-efficient segmentation of fraction MRIs with high geometrical accuracy and accurately segmented images obtained at different field strengths.

Objective. Uncertainty assessment of deep learning autosegmentation (DLAS) models can support contour corrections in adaptive radiotherapy (ART), e.g. by utilizing Monte Carlo Dropout (MCD) uncertainty maps. However, poorly calibrated uncertainties at the patient level often render these clinically nonviable. We evaluated population-based and patient-specific DLAS accuracy and uncertainty calibration and propose a patient-specific post-training uncertainty calibration method for DLAS in ART. Approach. The study included 122 lung cancer patients treated with a low-field MR-linac (80/19/23 training/validation/test cases). Ten single-label 3D-U-Net population-based baseline models (BM) were trained with dropout using planning MRIs (pMRIs) and contours for nine organs-at-riks (OARs) and gross tumor volumes (GTVs). Patient-specific models (PS) were created by fine-tuning BMs with each test patient’s pMRI. Model uncertainty was assessed with MCD, averaged into probability maps. Uncertainty calibration was evaluated with reliability diagrams and expected calibration error (ECE). A proposed post-training calibration method rescaled MCD probabilities for fraction images in BM (calBM) and PS (calPS) after fitting reliability diagrams from pMRIs. All models were evaluated on fraction images using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95) and ECE. Metrics were compared among models for all OARs combined (n = 163), and the GTV (n = 23), using Friedman and posthoc-Nemenyi tests (α = 0.05). Main results. For the OARs, patient-specific fine-tuning significantly (p < 0.001) increased median DSC from 0.78 (BM) to 0.86 (PS) and reduced HD95 from 14 mm (BM) to 6.0 mm (PS). Uncertainty calibration achieved substantial reductions in ECE, from 0.25 (BM) to 0.091 (calBM) and 0.22 (PS) to 0.11 (calPS) (p < 0.001), without significantly affecting DSC or HD95 (p > 0.05). For the GTV, BM performance was poor (DSC = 0.05) but significantly (p < 0.001) improved with PS training (DSC = 0.75) while uncertainty calibration reduced ECE from 0.22 (PS) to 0.15 (calPS) (p = 0.45). Significance. Post-training uncertainty calibration yields geometrically accurate DLAS models with well-calibrated uncertainty estimates, crucial for ART applications.

In multicentric studies, data sharing between institutions might negatively impact patient privacy or data security. An alternative is federated analysis by secure multiparty computation. This pilot study demonstrates an architecture and implementation addressing both technical challenges and legal difficulties in the particularly demanding setting of clinical research on cancer patients within the strict European regulation on patient privacy and data protection: 24 patients from LMU University Hospital in Munich, Germany, and 24 patients from Policlinico Universitario Fondazione Agostino Gemelli, Rome, Italy, were treated for adrenal gland metastasis with typically 40 Gy in 3 or 5 fractions of online-adaptive radiotherapy guided by real-time MR. High local control (21% complete remission, 27% partial remission, 40% stable disease) and low toxicity (73% reporting no toxicity) were observed. Median overall survival was 19 months. Federated analysis was found to improve clinical science through privacy-friendly evaluation of patient data in the European health data space.