Few-Shot Learning for Prostate Cancer Detection on MRI: Comparative Analysis with Radiologists' Performance

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning for prostate cancer detection requires large datasets, often hindered by domain shift issues across institutions.
• Limited data availability restricts the clinical applicability of current deep learning models.

Purpose of the Study:

• To develop a few-shot learning deep learning model for prostate cancer detection on multiparametric MRI.
• To evaluate the diagnostic performance of this model against experienced radiologists.
• To address data scarcity and domain shift challenges in AI-driven cancer detection.

Main Methods:

• A 2D transformer model was trained on T2-weighted, diffusion-weighted, and apparent diffusion coefficient (ADC) map images from 99 biopsy-confirmed prostate cancer cases.
• The model utilized a few-shot learning approach with minimal training data (20 cases).
• Performance was assessed using Matthews correlation coefficient (MCC) and F1 score, compared against two radiologists, with external validation on the Prostate158 dataset.

Main Results:

• The few-shot model achieved an MCC of 0.297 and an F1 score of 0.707, comparable to Radiologist 1 (MCC: 0.276, F1: 0.741).
• Radiologist 2 outperformed the model (MCC: 0.504, F1: 0.871).
• ImageNet pretraining significantly improved model performance, increasing study-level ROC-AUC from 0.464 to 0.636 and PR-AUC from 0.637 to 0.773.

Conclusions:

• Few-shot deep learning models, particularly with pretrained transformer architectures, can achieve clinically relevant performance for prostate cancer detection.
• This approach offers a viable solution to overcome data limitations and domain shift issues in multi-institutional settings.
• The findings support the potential of AI to enhance diagnostic capabilities in prostate cancer imaging.