Self-supervised learning leads to improved performance in biparametric prostate MRI classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Prostate cancer diagnosis relies heavily on volumetric imaging, particularly bi-parametric MRI (bpMRI).
• Current supervised learning models require extensive annotated data, which is labor-intensive to acquire.
• Self-supervised learning (SSL) offers a promising alternative for leveraging large unlabeled medical datasets.

Purpose of the Study:

• To develop and evaluate 2D self-supervised learning (SSL) models for volumetric MRI analysis.
• To demonstrate the efficacy of these SSL models in prostate cancer classification tasks using bpMRI.
• To compare the performance of SSL models against fully supervised learning (FSL) baselines.

Main Methods:

• Two distinct 2D SSL methods were trained on a large dataset of prostate multiparametric MRI (mpMRI) from 12 European centers.
• The pre-trained SSL models were transferred to volumetric prostate bpMRI classification tasks using attention-based multiple instance learning (MIL).
• Performance was evaluated across three tasks: prostate cancer diagnosis, clinically significant prostate cancer diagnosis, and virtual biopsy, using AUC and cross-validation.

Main Results:

• SSL models demonstrated comparable or superior performance to FSL baselines in several prostate cancer classification tasks.
• For bpMRI D-PCa, AUC was 0.82 for SSL vs. 0.75 for FSL (p=0.017); for T2 D-csPCa, AUC was 0.73 for SSL vs. 0.68 for FSL (p=0.043).
• SSL models required less training data to achieve similar performance, and attention scores correlated with lesion location.

Conclusions:

• Unsupervised SSL models trained on unlabeled data are more data-efficient and perform better than FSL models in volumetric prostate MRI classification.
• These findings underscore the critical importance of large-scale data collection and annotation efforts in advancing biomedical imaging AI.
• SSL presents a viable strategy to overcome data limitations in medical image analysis.