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

Deep-learning models for prostate cancer detection often require large datasets, which can be challenging to obtain and may lead to domain shift issues in various clinical settings. This study aimed to develop a deep-learning model for prostate cancer detection on magnetic resonance images using few-shot learning and compare its performance with radiologists. This retrospective study used 99 cases (80 positive, 19 negative) of confirmed prostate cancer, diagnosed through needle biopsy from 2017 to 2022, with 20 cases for training, 5 for validation, and 74 for testing. The 2D transformer model was trained on T2-weighted, diffusion-weighted, and apparent diffusion coefficient map images. Model predictions were compared between the two radiologists using the Matthews correlation coefficient (MCC) and F1 score, and the bootstrap method was used to calculate 95% confidence intervals (CIs). Seventy-four patients (mean age, 71 years ± 8; 60 men) were included in the test set. The model achieved an MCC of 0.297 (95% CI: 0.095–0.474) and F1 score of 0.707 (95% CI: 0.598–0.847). Radiologist 1 had an MCC of 0.276 (95% CI: 0.054–0.484) and an F1 score of 0.741 (95% CI: 0.632–0.832), while Radiologist 2 had an MCC of 0.504 (95% CI: 0.289–0.703) and an F1 score of 0.871 (95% CI: 0.800–0.931). The performance of the model was not significantly different from that of Radiologist 1 (MCC difference: 0.021, 95% CI: −0.270–0.306; F1 score difference: −0.034, 95% CI: −0.153–0.078), but was lower than that of Radiologist 2 (F1 difference: −0.16, 95% CI: −0.287– - 0.061). A deep-learning model trained on only 20 cases achieved a performance comparable to one radiologist in detecting prostate cancer on magnetic resonance images, demonstrating the potential of few-shot learning in addressing domain shift challenges. A deep learning model for prostate cancer detection on MRI was developed using only 20 training cases. The model achieved performance comparable to one radiologist (MCC: 0.297 vs 0.276) but lower than another (F1: 0.707 vs 0.871). Few-shot learning demonstrated potential for addressing domain shift challenges in medical imaging AI. Few-shot learning enables development of prostate cancer detection models on MRI with performance comparable to radiologists, using minimal training data.