Efficient Deep Ladle-Net for fast universal 3D lesion segmentation on chest-abdomen-pelvis computed tomography

Accurate evaluation of tumor size on follow-up computed tomography (CT) scans is critical for assessing treatment efficacy in cancer patients. However, universal lesion segmentation of CT remains a significant challenge due to the complexity and diversity of chest-abdomen-pelvis imaging. Traditional manual segmentation is time-consuming, labor-intensive, and subject to inter-observer variability due to its reliance on radiologists' expertise. Current AI models are often limited to segmentation of a single type of lesion, lacking the capability to handle multiple lesion types within a single framework. Moreover, the high computational cost to analyze volumetric CT data and model the intricate and diverse lesion patterns makes universal lesion segmentation even harder. To address the above-mentioned issues, we introduce an effective and efficient Deep Ladle-Net for fast universal 3D lesion segmentation on the chest-abdomen-pelvis CT scans to examine 10 types of lesions at once, including abdominal lesions, bone lesions, colon lesions, kidney lesions, liver lesions, lung lesions, mediastinal lesions, pancreas lesions, lung nodules, and lymph node lesions. Our framework was evaluated on a large cohort with 7151 lesions from 11 publicly available 3D CT datasets and two private datasets. The experimental results show that our preliminary model achieved excellent performance, outperforming five state-of-the-art methods and ranking third in the 2024 Universal Lesion Segmentation Challenge (ULS23). With improved data normalization and augmentation, the proposed Deep Ladle-Net obtains overall Segmentation Dice 0.773±0.146 on 10 lesion types in our extended study and is particularly good in abdominal lesions, kidney lesions, mediastinal lesions and lung nodules, achieving Segmentation Dice greater than 0.78. Run-time analysis shows that the proposed framework achieves high computational efficiency and takes less than 2 s per case using a single NVIDIA Geforce RTX4080 GPU card. These findings highlight the potential of the proposed method to streamline radiological workflows, support early detection, and improve clinical decision-making with reliable and reproducible 3D lesion segmentation results.