Automatic segmentation of liver tumors based on attention-guided 3D U-Net using LiTS dataset revision
4 affiliations
- 1Department of Radiology & Imaging Center, Hunan Cancer Hospital, Hunan, 410000, China
- 2Department of Cardiology, Hunan Provincial People's Hospital, China
- 3Endoscopy Diagnosis and Treatment Center, Hunan Cancer Hospital, Hunan, 410000, China
- 4Endoscopy Diagnosis and Treatment Center, Hunan Cancer Hospital, Hunan, 410000, China; Corresponding author.
Abstract
Objective: To develop and evaluate an attention-guided 3D U-Net framework for liver tumor segmentation on contrast-enhanced CT images, with particular emphasis on improving segmentation performance under severe foreground-background imbalance caused by the sparse distribution of tumor regions. Methods: The 131 labeled LiTS cases were divided by patient into 92 training cases, 20 validation cases, and 19 held-out internal test cases. CT volumes were clipped to −160 to 240 HU, normalized to [0, 1], and cropped into 64 × 64 × 64 patches. Safe 3D augmentation was applied only to the training set. Attention gates were embedded in the skip connections of a 3D U-Net, and a hybrid Dice–binary cross-entropy loss was used to address foreground–background imbalance. The proposed model was compared with eight representative 3D segmentation networks using Dice, intersection over union (IoU), precision, and recall. Results: On the independent test set, on the held-out internal test subset of 19 LiTS patients, the proposed model achieved a Dice score of 0.841 (95% CI: 0.812–0.870) and IoU of 0.765 (95% CI: 0.731–0.799), outperforming all evaluated baseline models. Compared with the locally implemented vanilla 3D U-Net, Dice increased by 0.1401 (14.01 percentage points; 19.98% relative) and IoU increased by 0.1642 (16.42 percentage points; 27.34% relative). Qualitative examples suggested improved boundary continuity in selected difficult cases; this observation was not a prespecified subgroup analysis. Conclusion: The proposed framework improved liver tumor segmentation performance on the LiTS dataset under a controlled evaluation protocol. Further external multicenter validation is required to assess its generalizability across different clinical settings. Further multicenter validation is required before clinical application.
Related
- Semi-automatic segmentation of pelvic bone tumors: Usability testing
- Deep co-supervision and attention fusion strategy for automatic COVID-19 lung infection segmentation on CT images
- Fully automatic airway segmentation in CT images via dual-stage neural network with topology-constrained centerline completion
- Attention-Deficit/Hyperactivity Disorder Is Associated With Lower Rates of Adhesive Capsulitis, Bursitis, Shoulder Stiffness, and Revision Shoulder Arthroscopy After Arthroscopic Rotator Cuff Repair
- Segmentation and morphological analysis of recanalized coiled aneurysms: Towards an automatic tool for clipping planning in microsurgery
BibTeX
@article{gong2026automatic,
title = {{Automatic segmentation of liver tumors based on attention-guided 3D U-Net using LiTS dataset revision}},
author = {Min Gong and Jiaru Wu and Ang Qin and Yong Qiu},
journal = {Journal of Radiation Research and Applied Sciences},
year = {2026},
month = oct,
doi = {10.1016/j.jrras.2026.102717},
url = {https://doi.org/10.1016/j.jrras.2026.102717},
note = {CC BY},
}
RIS
TY - JOUR
TI - Automatic segmentation of liver tumors based on attention-guided 3D U-Net using LiTS dataset revision
AU - Min Gong
AU - Jiaru Wu
AU - Ang Qin
AU - Yong Qiu
PY - 2026
DA - 2026/10/30
JO - Journal of Radiation Research and Applied Sciences
DO - 10.1016/j.jrras.2026.102717
UR - https://doi.org/10.1016/j.jrras.2026.102717
AB - Objective: To develop and evaluate an attention-guided 3D U-Net framework for liver tumor segmentation on contrast-enhanced CT images, with particular emphasis on improving segmentation performance under severe foreground-background imbalance caused by the sparse distribution of tumor regions. Methods: The 131 labeled LiTS cases were divided by patient into 92 training cases, 20 validation cases, and 19 held-out internal test cases. CT volumes were clipped to −160 to 240 HU, normalized to [0, 1], and cropped into 64 × 64 × 64 patches. Safe 3D augmentation was applied only to the training set. Attention gates were embedded in the skip connections of a 3D U-Net, and a hybrid Dice–binary cross-entropy loss was used to address foreground–background imbalance. The proposed model was compared with eight representative 3D segmentation networks using Dice, intersection over union (IoU), precision, and recall. Results: On the independent test set, on the held-out internal test subset of 19 LiTS patients, the proposed model achieved a Dice score of 0.841 (95% CI: 0.812–0.870) and IoU of 0.765 (95% CI: 0.731–0.799), outperforming all evaluated baseline models. Compared with the locally implemented vanilla 3D U-Net, Dice increased by 0.1401 (14.01 percentage points; 19.98% relative) and IoU increased by 0.1642 (16.42 percentage points; 27.34% relative). Qualitative examples suggested improved boundary continuity in selected difficult cases; this observation was not a prespecified subgroup analysis. Conclusion: The proposed framework improved liver tumor segmentation performance on the LiTS dataset under a controlled evaluation protocol. Further external multicenter validation is required to assess its generalizability across different clinical settings. Further multicenter validation is required before clinical application.
SN - 2314-7164
SN - 1687-8507
ER -