Semi-Supervised Learning in Prostate MRI Tumor Segmentation Approaches Fully-Supervised Performance on External Validation
Eduardo Pooch, Georgios Agrotis, Lishan Cai, Mark Emberton, Taimur Tariq Shah, Hashim Uddin Ahmed, Regina G. H. Beets‐Tan, S. Benson, Tomas M. Janssen, Ivo G. Schoots · medRxiv · 2025
Abstract Purpose To evaluate the diagnostic performance of semi-supervised learning models for aggressive prostate cancer segmentation on MRI compared to fully-supervised models trained with additional expert annotations. Materials and Methods We used 1500 MRI scans from the PI-CAI challenge training subset. Positive scans had 220 human and 205 AI-generated annotations. The mtU-Net (proposed teacher-student semi-supervised approach) was compared to supervised (trained using only 220 human annotations) and semi-supervised (trained on human and AI-generated annotations) nnU-Net. The 205 AI-annotated scans were manually annotated, and a fully-supervised model was trained. External validation was performed on a newly annotated dataset from the PROMIS study (n=574) and the Prostate158 dataset (n=158). Patient-level performance was evaluated using Area Under the Curve (AUC), Average Precision (AP) for lesion-level detection, and the DeLong test to compare performance. Results The fully-supervised nnU-Net showed the highest performance on the internal PI-CAI test set (AUC=0.89[0.87-0.91]/AP=0.65[0.60-0.70]) and external validation datasets PROMIS (AUC=0.70[0.66-0.74]/AP=0.24[0.19-0.29]) and Prostate158 (AUC=0.87[0.82-0.92]/AP=0.64[0.56-0.72]), significantly outperforming the supervised baseline (p≤0.002). The proposed semi-supervised mtU-Net demonstrated close external validation performance on PROMIS (AUC=0.66[0.62-0.71]/AP=0.20[0.16-0.25]) and Prostate158 (AUC=0.86[0.81-0.92]/AP=0.58[0.49-0.67]), significantly outperforming the supervised baseline on both datasets (p=0.024 and p=0.007, respectively). Semi-supervised nnU-Net showed intermediate results on PROMIS (AUC=0.65[0.60-0.69]/AP=0.20[0.16-0.24]) and Prostate158 (AUC=0.81[0.74-0.88]/AP=0.53[0.44-0.62]), significantly outperforming the supervised baseline only on PROMIS (p=0.042). Conclusion In prostate MRI tumor segmentation, nnU-Net fully-supervised learning performed best. However, in external validation, mtU-Net’s semi-supervised learning performance approached the fully-supervised model, demonstrating a valuable approach when expert annotations are limited. Summary Semi-supervised learning achieves close performance to fully-supervised methods on external validation in prostate cancer segmentation, reducing dependence on expert annotations in increasing demands. Key points The inclusion of AI-annotated data during training showed close performance to annotating additional samples with expert delineations, suggesting data diversity may be as impactful as increased expert annotation volume. The combination of pseudo-labeling with consistency regularization within the semi-supervised mtU-Net framework mitigated the impact of potential inaccuracies in AI-generated annotations, resulting in performance approaching that of fully-supervised models.