582 Generalist vs specialist models for medical image segmentation

Dr. Andrea Moglia

Accurate reconstruction of radiological volumes through segmentation plays a crucial role in understanding patient anatomy, supporting experts in surgical planning, especially in laparoscopy and robotic surgery, where the restricted visualized area during surgery makes it essential to clearly recognize the patient anatomy. In the past decade there has been a tremendous growth of deep learning specialist models for segmentation. More recently, numerous generalist models, pre-trained on millions of multi-modality images across diverse anatomical regions, have been developed, following the paradigm shift of large language models. We performed the first survey on the published literature on generalist models for medical image segmentation, with a comparison with state-of-the-art specialist models on 15 anatomical areas on 58 public datasets. Generalist models obtained the highest Dice score in all anatomical sites, with the exception of thoracic structures. To support further research, we created the first database-backed web application (https://hal9000-lab.github.io/GMMIS-Survey/).

See more at: https://vattikutifoundation.com/videos/