Author: Ahmad Algohary, Adrian Breto, Quadre Emery, Radka Stoyanova 👨🔬
Affiliation: University of Miami, Department of Radiation Oncology, University of Miami 🌍
Purpose:
To develop a foundation model (U-Found) for multiparametric MRI (mpMRI) of the prostate by using self-supervised learning to prove the feasibility of a prostate-oriented foundation model using a large publicly available dataset.
Methods:
The PI-CAI Grand Challenge’s public dataset were used to train the embedding encoders. Each case consisted of T2-weighted (T2w) imaging, apparent diffusion coefficient maps (ADC), and the high b-value diffusion-weighted imaging (HBV) from either a Siemens or Philips magnet. The volume of the tumor on each 2D axial slice, together with the lesion Grade Group (GG), was available. The encoder was developed on the SimCLR contrastive learning schema with ResNet-50 as a backbone network.
Results:
Four networks were trained, using a single sequence (T2w, ADC or HBV) or combined, using 1500 cases (~22k axial slices). The embeddings (128 dim) were reviewed using the Unifrom Manifold Approximation and Projection (UMAP) algorithm and the ADC maps demonstrated clusters with high tumor content. 11 deep features were selected by minimal redundance, maximum relevance (MRMR), discriminating for cancer. We carried logistic regression analysis, first using the embeddings from all the data and second – only from representatives of the “tumor” cluster. The high cancer content of this cluster is confirmed by a higher AUC (0.98 vs. 0.94).
Conclusion:
The high concentration of visible cancer in certain areas of the embedding manifold demonstrates the high effectiveness of U-Found in assessing features of prostate mpMRI. This result is remarkable not only because U-Found can learn features of cancer without labels or supervision but also because of the high efficacy of U-Found in creating important representations. The presence of meaningful features in the outputs of the network proves that the embeddings contain important information. In conclusion, we consider U-Found a large step towards a foundation model for prostate MRI.