ABSTRACT Accurate 3D lung tumor segmentation on CT is crucial for stereoscopic diagnosis and virtual preoperative planning. However, this task faces significant challenges: traditional segmentation methods are often limited by parameter sensitivity and manual intervention, while purely deep learning approaches are constrained by the scarcity of annotated medical data and struggle with irregular, low‐contrast tumor margins. To overcome these limitations, we propose a fully automatic hybrid pipeline that combines AI‐based localization with an interpretable segmentation algorithm. The method first uses a pretrained MONAI 3D RetinaNet detector to localize tumors and generate seeds, followed by a four‐point ensemble region growing process with adaptive thresholding and dynamic constraints, and finally reconstructs the tumor using the Marching Cubes algorithm. Evaluated on selected cases from the LIDC‐IDRI dataset, the proposed method achieved a mean Dice score of 0.935 for tumor segmentation, exceeding other models like I‐3D DenseUNet (0.831), 3D MultiResUNet (0.866), and a hybrid CNN (0.720). By synergizing deep learning for robust localization and a constrained region‐growing algorithm for precise boundary delineation, our method provides stable 3D tumor contours and interactive models, showing great promise for assisting thoracoscopic virtual preoperative planning.
Li et al. (Sun,) studied this question.