Localization of Non-Isocentric Wrist Anatomy with Few-shot Foundation Models

Motivation: Automatic localization of non-isocentric anatomy such as wrist for autolocalization workflow is challenging due to its flexible scan positions and field of views. In this abstract, we develop and compare capability of few-shot foundation model-based localization of wrist. Goal(s): Develop and compare few-shot foundation models (FM) localization capability for wrist anatomy. Approach: Self-supervised adapter, and convolution neural network (CNN) based inference time few-shot task adaptation models are developed and compared for 3D wrist localization capability. Results: Self-supervised few-shot adapter-based model found to be superior for 3D localization with significantly high localization accuracy and less false positives and false negatives compared to CNN approach. Impact: With as few as 6 labeled images, localization adapter trained with self-supervised visual features performed significantly better than CNN based few-shot model that used as many as 28 image-mask pairs. Demonstrated 3D localization capability from models trained on 2D images.