Tumor-SAM: Segment Anything Model for Semi-automatic Lung Tumor Segmentation in CT

Lung tumor segmentation in thoracic CT scans is vital for radiomics analysis and treatment assessment, but is hindered by heterogeneous tumor morphology, ambiguous boundaries, and inaccuracy and labor-intensiveness of manual segmentation. Existing methods, including traditional machine learning and deep learning methods, often suffer from over-/under-segmentation or poor robustness. In this study, we propose an improved Segment Anything Model (called Tumor-SAM) for semi-automatic lung tumor segmentation, integrating U-Net for multi-scale feature extraction and a novel ellipse prompt. Tumor-SAM first detects lung ROI to reduce interference from the surrounding tissue. Then, we design an ellipse prompt defined by center, axes, and rotation that captures tumor shape/location better than points, boxes, or circles. The architecture of Tumor-SAM includes a U-Net-based image encoder (replacing ViT), prompt encoder with positional encoding, multi-head attention fusion, and mask decoder. Our method achieved an average Dice index of 0.84±0.13 and an average Hausdorff distance of 7.25±6.24 mm on 164 testing scans, demonstrating good lung tumor segmentation accuracy.