Beyond geometric accuracy: Evaluation of an MRI-based autocontouring prototype for brain and prostate radiotherapy

Abstract

Abstract Background MRI-based autocontouring has the potential to improve radiotherapy planning and adaptive workflows by reducing contouring time while maintaining clinical quality. We evaluated a prototype deep learning–based MRI autocontouring tool for organ-of-interest (OOI) and target delineation in brain metastases and prostate cancer. Methods Fifty-seven patients (27 brain metastases and 30 prostate cancer) were retrospectively analysed. Automatically generated contours were compared with clinically approved reference contours using the Dice Similarity Coefficient (DSC) and Mean Distance to Agreement (MDA). Two radiation oncologists independently assessed clinical usability using a four-point scoring system. In addition, target detection sensitivity and false-prediction rate were evaluated for brain metastases. Results A total of 988 structures were generated. In the brain cohort, the prototype achieved a target detection sensitivity of 81.5% with a false-prediction rate of 20.9%. Mean DSC was 0.58 for OOI and 0.66 for gross tumour volumes, while median clinical usability was 3 (minor editing required). In the prostate cohort, mean DSC was 0.73 and median clinical usability was also 3. Performance varied substantially across structures, with large, well-defined anatomy showing the highest accuracy, whereas small or anatomically complex structures such as the cochlea, and seminal vesicles required more extensive correction. Lower DSC values were significantly associated with poorer clinical acceptability for 10 of 19 evaluated structures. Conclusions The prototype demonstrated promising performance as a clinician-in-the-loop contouring aid for MRI-guided radiotherapy. Although it can reduce manual contouring workload, expert review remains essential, particularly for target delineation and small or anatomically complex structures. A combined evaluation using geometric, detection, and clinical metrics provides a more comprehensive assessment than geometric measures alone.