Zero-shot swift lung 4D CT segmentation via optimized Segment Anything Model

Segmenting lung CT images is crucial for radiation oncology, impacting treatment planning, tumor tracking, and response assessment. This study presents an optimized version of the Segmentation Anything Model (SAM), known as RepViT-SAM, tailored for efficient and accurate lung CT segmentation. RepViT-SAM incorporates the RepViT encoder, which enhances computational efficiency while maintaining high segmentation performance. The model utilizes minimal user input through point prompts to perform zero-shot segmentation, adapting to real-time clinical requirements. We evaluated RepViT-SAM on a 4D lung CT dataset, achieving an average Dice Similarity Coefficient (DSC) of 87.9% and a pixel accuracy of 96.8%, with a significantly reduced latency of 32.7ms—outpacing both the original SAM and MobileSAM models. These results demonstrate that RepViT-SAM is not only highly accurate but also fast enough to meet the demands of clinical applications, including real-time lung CT segmentation. Future work will focus on expanding the model's capabilities to 3D segmentation and improving its robustness across varying lung morphologies.