Context-Aware Multi-Organ Segmentation in Abdominal CT via LoRA-Fine-tuned MedSAM

Automated, accurate and robust segmentation of CT scans on abdominal organs remains a formidable challenge due to the blurred boundaries and subtle features of adjacent structures of abdominal organs. MedSAM, the fully fine-tuned model based on the Segment Anything Model (SAM) for medical image, fails to perform well relying solely on box prompt. Besides, utilizing the contextual information in the 3D volume data is tricky. To address above issues, we develop a MedSAM-based model for abdominal multi-organ CT segmentation. A feature selection network is introduced to extract entropy and curvature information. Together with the image embedding from the image encoder of MedSAM, the extracted information is processed during the inference. Low-Rank Adaptation (LoRA) is also applied to the image encoder and mask decoder of MedSAM for efficient fine-tuning on the FLARE2022 dataset. During the inference, mask prompts from segmentation results of last adjacent structure together with the expanded box prompt are used to improve the performance in continued context, thereby achieving an automated workflow. The experimental results suggest that the proposed method demonstrates a state-of-the-art performance and a potential clinical promise.