Kidney and Kidney Tumor Segmentation via Transfer Learning

Recently Segment anything model (SAM) has shown great promise for natural image segmentation. This model was trained on by far the largest segmentation dataset, consisting of over 11 million diverse images and 1 billion corresponding masks. The dataset’s impressive size and high quality, combined with the powerful Transformer-based architecture, enabled the model to grasp a general understanding of objects and achieve exceptional zero-shot performances, sometimes even outperforming fully supervised models. However, despite the significant advancements within the zero-shot framework, there are challenges when applying it to more specialized domains like medical and satellite imaging. Due to the scarcity of images from those domains in the training corpus, the model is not as accurate as it could be. Additionally in the fields where the segmentation of only certain, critical areas is desired using the SAM model can be overwhelming. In this paper, We aim to make use of different Transfer Learning techniques, such as Feature Extraction and Fine-tuning, and investigate different slight adaptations of the architecture to improve the performance of the SAM model and achieve high performance on a given medical image segmentation task.