Volumetric CT Segmentation with Mask Propagation Using Segment Anything
Huu-Hung Nguyen, Cong-Tai Nguyen, Minh–Triet Tran · 2023
Medical imaging is both an interesting and challenging field for researchers, mainly due to their lack of labeled data, especially in segmentation tasks. Many interactive system has been introduced to help streamline the annotation workflow, and most recently Segment Anything Model (SAM) has been a breakthrough as a foundation model in interactive segmentation using prompts. In this paper, we adapt SAM, a 2D segmentation for 3D organ interactive segmentation task, then we propose several performance improvement strategies and achieve a comparable result of 0.8694 of mean DSC in a full-supervised setting with a small amount of data. Furthermore, in the process of finding a method to reduce the human effort when using our algorithm, we also develop a novel method inspired by a beam search algorithm, stemming from the NLP domain, which can run volumetric segmentation for one target (i.e. the liver) with minimal human manual input, and still achieve an interesting result of 0.9410 Liver’s DSC, allow it to be integrated to annotation system to aid medical expert.