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

BACKGROUND AND

OBJECTIVE

Develop two-dimensional self-supervised learning (SSL) models which can be used in volumetric imaging and demonstrate their application in volumetric prostate bi-parametric MRI (bpMRI) classification tasks.

METHODS

Prostate multiparametric MRI (mpMRI) data from 12 distinct European centers were used to train two SSL methods. We transfer these models to classification tasks in volumetric prostate bpMRI using 3 attention-based multiple instance learning (MIL) methods with T2-weighted (T2) or bpMRI studies. Three prostate cancer (PCa) tasks were considered: PCa diagnosis (D-PCa), clinically significant PCa (csPCa) diagnosis (D-csPCa), and virtual biopsy to confirm csPCa (VB). All approaches were compared with a fully supervised learning (FSL) baseline. Performance was assessed using the area under the receiver operating curve (AUC) and using both 5-fold cross-validation and a hold-out test set, and attention scores were analyzed. Finally, sensitivity analyses were performed for training and pre-training dataset size, data domain (MRI vs. natural images), and architecture.

RESULTS

Two 2D SSL methods were trained using 6,798 studies (1,722,978 DICOM images) and their downstream performance was assessed on 3D tasks (n=1,622, n=1,615 and n=1,295 bmMRI studies for D-PCa, D-csPCa and VB, respectively). We show these models are comparable or better than FSL baseline models trained on the same data: AUCSSL=0.82 and AUCFSL=0.75 for bpMRI D-PCa (p=0.017), AUCSSL=0.73 and AUCFSL=0.68 for T2 D-csPCa (p=0.043) and AUCSSL=0.73 and AUCFSL=0.65 for bpMRI VB, while other models showed no differences (p>0.05). Learning curve analyses show that SSL-based models required fewer training data to perform similarly, while sensitivity analyses showed that large amounts of domain-specific pre-training data are essential for optimal performance. Attention scores correlate with lesion location.

CONCLUSION

Data with no annotations was used to train SSL models which were more data efficient and performed better than FSL models, highlighting the importance of large-scale data collection efforts in biomedical imaging.