DINO ‐ SAM : Combining DINO Detector and Fine‐Tuned SAM for Automatic Pulmonary Nodule Segmentation in CT Image

ABSTRACT Accurate pulmonary nodule segmentation in computed tomography (CT) images is important for the quantitative assessment of nodule morphology. However, the small size, low contrast, and heterogeneous appearance of pulmonary nodules make it difficult to delineate their boundaries accurately. Although the Segment Anything Model (SAM) has demonstrated strong general‐purpose segmentation capability, its dependence on manual prompts and limited adaptability to pulmonary CT images hinder fully automatic segmentation. To address these challenges, we propose DINO‐SAM, a fully automatic pulmonary nodule segmentation framework that combines a DINO detector with a parameter‐efficiently fine‐tuned SAM. The DINO detector automatically localizes pulmonary nodules and generates prompts for SAM, thereby eliminating manual interaction. Low‐Rank Adaptation (LoRA) is incorporated into the SAM image encoder and mask decoder, while position and feature adapters are introduced into the image encoder to enhance pulmonary‐nodule‐specific representation learning with a limited number of trainable parameters. DINO‐SAM was evaluated on the public LIDC‐IDRI dataset and achieved a Dice of 92.42%, an IoU of 86.11%, an HD of 2.44 pixels, an ACC of 99.52%, an SE of 93.37%, and an SP of 99.72%. The proposed method outperformed the compared task‐specific and SAM‐based foundation models. These results demonstrate that detector‐guided prompt generation and parameter‐efficient SAM adaptation provide an effective approach to fully automatic pulmonary nodule segmentation in CT images.