Self-supervised Learning Network on Large Prostate Cancer mpMRI Dataset: Towards A Foundational Model of the Prostate
Noah C. Lowry, Adrian Breto, Veronica Wallaengen, Ahmad Algohary, Radka S. Stoyanova · Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025
Motivation: No foundation models currently exist for prostatic mpMRI analysis. Goal(s): To develop a foundation model (U-Found) and to evaluate its embeddings for a series of downstream tasks. Approach: Development of an encoder neural network that learns vector representations of prostate mpMRI through contrastive learning. Results: U-Found embeddings successfully encode features of prostate MRI including presence of cancer without ever explicitly learning those labels under the self-supervised framework. Impact: To the best of our knowledge, U-Found is the first foundation-like model developed for prostate mpMRI. The embeddings, combining cancer and overall prostate characteristics features can be used in comprehensive modeling of cancer progression or response to therapy.