Few-Shot Contrastive Multilabel Localization of Knee MR Images with Self-Supervised Foundation Model

Motivation: Develop automatic multilabel localization capability on knee MRI images with minimal manual annotation Goal(s): Simultaneous multilabel localization of knee images to reduce the annotation time by leveraging few labelled images Approach: Trained a contrastive multilabel adopter from the features derived from the self-supervised MR vision foundation model (FM) to simultaneously localize multiple regions of interest (RoI) on the MR knee images Results: Excellent knee simultaneous RoI localization (88%, 88% and 77%) on three anatomies with FM with only few labelled images Impact: Multilabel contrastive model trained by features extracted from FM with few labeled examples showed promising results for localizing multiple RoI's on knee images. Multilabel model showed excellent localization across knee slices by preventing false positives.