Dissecting self-supervised learning strategies for transfer learning in MRI prostate cancer diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning (DL) shows promise in medical imaging but struggles with limited annotated data and generalization to out-of-distribution (OOD) datasets.
• Self-supervised learning (SSL) and transfer learning offer solutions by leveraging unlabeled data and adapting models to new domains.

Purpose of the Study:

• To evaluate SSL-based strategies for detecting and segmenting clinically significant prostate cancer (csPCa) in MRI.
• To investigate the impact of model architectures, pretext tasks, contrastive learning, and downstream tasks on performance.
• To assess the effectiveness of a three-stage training pipeline (SSL, supervised pre-training, fine-tuning) on diverse datasets.

Main Methods:

• Utilized a medium-sized pre-training dataset (PI-CAI) and two small OOD target datasets (Prostate158, ChiPCa).
• Implemented a three-stage training pipeline involving SSL, supervised pre-training, and fine-tuning.
• Evaluated various SSL strategies, model architectures (UNETR, UNet), and dataset similarities.

Main Results:

• The full three-stage pipeline showed consistent performance for csPCa detection and segmentation on Prostate158 (similar distribution to pre-training data).
• For ChiPCa (different distribution), the full pipeline excelled in detection but was suboptimal for segmentation, where partial training stages performed better.
• UNet architecture generally yielded superior segmentation results compared to UNETR.

Conclusions:

• Multi-stage SSL pipelines are beneficial, but their optimal configuration depends on dataset similarity to pre-training data.
• Model architecture choice (UNet for segmentation) and dataset characteristics significantly impact performance in csPCa detection and segmentation.
• Findings offer practical guidance for applying SSL in medical imaging for prostate cancer analysis.