Patient-specific uncertainty calibration of deep learning-based autosegmentation networks for adaptive MRI-guided lung radiotherapy.
Authors
Affiliations (7)
Affiliations (7)
- Radiation Oncology, Klinikum der Universität München, Marchioninistr. 15, Munchen, 80377, GERMANY.
- Department of Radiotherapy, University Medical Centre Utrecht, Heidelberglaan 100, Utrecht, 3584 CX, NETHERLANDS.
- Department of Radiation Oncology, LMU University Hospital, Marchioninistr. 15, LMU Munich, 81377, GERMANY.
- Department of Radiation Oncology, University Hospital of Ludwig-Maximilians-University Munich, Marchioninistr. 15, Munich, 81377, GERMANY.
- Radiation Oncology, University Hospital Munich Campus Grosshadern, Marchioninistr. 15, Munchen, 81377, GERMANY.
- Department of Radiotherapy, University Medical Centre Utrecht, Room Q00.118, Heidelberglaan 100, 3584 CX Utrecht, THE NETHERLANDS, Utrecht, 3508 GA, NETHERLANDS.
- Radiation Oncology, University Hospital Munich, Marchioninistr. 15, Munich, Bavaria, 81377, GERMANY.
Abstract
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 14mm (BM) to 6.0mm (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.