Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MSWAL: 3D Multi-class Segmentation of Whole Abdominal Lesions Dataset

Zhaodong Wu, Qiaochu Zhao, Ming Hu, Yulong Li, Haochen Xue, Kang Dang, Zhengyong Jiang, Angelos Stefanidis, Qiufeng Wang, Imran Razzak, Zongyuan Ge, Junjun He, Yu Qiao, Zhong Zheng, Feilong Tang, Jionglong Su

TL;DR

This work introduces MSWAL, the first large-scale 3D multi-class segmentation dataset for whole-abdominal lesions, covering seven lesion types across 694 CT scans and ensuring full, consistent annotations. It demonstrates strong generalization via transfer learning to LiTS and KiTS and presents Inception nnU-Net, a hybrid architecture that fuses Inception modules with nnU-Net to capture multi-scale lesion information, achieving state-of-the-art performance on MSWAL. Extensive experiments, including ablations and a transfer-learning study, show substantial improvements in lesion segmentation and highlight remaining challenges such as cross-lesion interference and class-imbalance. The dataset and code provide a new benchmark and practical pathway toward more accurate, automated abdominal lesion diagnosis, with potential clinical impact in reducing radiologist workload and improving diagnostic specificity.

Abstract

With the significantly increasing incidence and prevalence of abdominal diseases, there is a need to embrace greater use of new innovations and technology for the diagnosis and treatment of patients. Although deep-learning methods have notably been developed to assist radiologists in diagnosing abdominal diseases, existing models have the restricted ability to segment common lesions in the abdomen due to missing annotations for typical abdominal pathologies in their training datasets. To address the limitation, we introduce MSWAL, the first 3D Multi-class Segmentation of the Whole Abdominal Lesions dataset, which broadens the coverage of various common lesion types, such as gallstones, kidney stones, liver tumors, kidney tumors, pancreatic cancer, liver cysts, and kidney cysts. With CT scans collected from 694 patients (191,417 slices) of different genders across various scanning phases, MSWAL demonstrates strong robustness and generalizability. The transfer learning experiment from MSWAL to two public datasets, LiTS and KiTS, effectively demonstrates consistent improvements, with Dice Similarity Coefficient (DSC) increase of 3.00% for liver tumors and 0.89% for kidney tumors, demonstrating that the comprehensive annotations and diverse lesion types in MSWAL facilitate effective learning across different domains and data distributions. Furthermore, we propose Inception nnU-Net, a novel segmentation framework that effectively integrates an Inception module with the nnU-Net architecture to extract information from different receptive fields, achieving significant enhancement in both voxel-level DSC and region-level F1 compared to the cutting-edge public algorithms on MSWAL. Our dataset will be released after being accepted, and the code is publicly released at https://github.com/tiuxuxsh76075/MSWAL-.

MSWAL: 3D Multi-class Segmentation of Whole Abdominal Lesions Dataset

TL;DR

This work introduces MSWAL, the first large-scale 3D multi-class segmentation dataset for whole-abdominal lesions, covering seven lesion types across 694 CT scans and ensuring full, consistent annotations. It demonstrates strong generalization via transfer learning to LiTS and KiTS and presents Inception nnU-Net, a hybrid architecture that fuses Inception modules with nnU-Net to capture multi-scale lesion information, achieving state-of-the-art performance on MSWAL. Extensive experiments, including ablations and a transfer-learning study, show substantial improvements in lesion segmentation and highlight remaining challenges such as cross-lesion interference and class-imbalance. The dataset and code provide a new benchmark and practical pathway toward more accurate, automated abdominal lesion diagnosis, with potential clinical impact in reducing radiologist workload and improving diagnostic specificity.

Abstract

With the significantly increasing incidence and prevalence of abdominal diseases, there is a need to embrace greater use of new innovations and technology for the diagnosis and treatment of patients. Although deep-learning methods have notably been developed to assist radiologists in diagnosing abdominal diseases, existing models have the restricted ability to segment common lesions in the abdomen due to missing annotations for typical abdominal pathologies in their training datasets. To address the limitation, we introduce MSWAL, the first 3D Multi-class Segmentation of the Whole Abdominal Lesions dataset, which broadens the coverage of various common lesion types, such as gallstones, kidney stones, liver tumors, kidney tumors, pancreatic cancer, liver cysts, and kidney cysts. With CT scans collected from 694 patients (191,417 slices) of different genders across various scanning phases, MSWAL demonstrates strong robustness and generalizability. The transfer learning experiment from MSWAL to two public datasets, LiTS and KiTS, effectively demonstrates consistent improvements, with Dice Similarity Coefficient (DSC) increase of 3.00% for liver tumors and 0.89% for kidney tumors, demonstrating that the comprehensive annotations and diverse lesion types in MSWAL facilitate effective learning across different domains and data distributions. Furthermore, we propose Inception nnU-Net, a novel segmentation framework that effectively integrates an Inception module with the nnU-Net architecture to extract information from different receptive fields, achieving significant enhancement in both voxel-level DSC and region-level F1 compared to the cutting-edge public algorithms on MSWAL. Our dataset will be released after being accepted, and the code is publicly released at https://github.com/tiuxuxsh76075/MSWAL-.

Paper Structure

This paper contains 13 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. MSWAL
  3. Dataset Statistical Analysis.
  4. Ground Truth Generation.
  5. Inception nnU-Net
  6. Mini Inception.
  7. Inception Downsampling.
  8. Experiments and Results
  9. Implementation Details.
  10. Comparative Experiment.
  11. Ablation Experiment.
  12. Transfer Learning Experiment.
  13. Conclusion

Figures (3)

  • Figure 1: Data distributions and an example of MSWAL. (a) gender distribution (male and female); (b) phase distribution (non-contrast, arterial phase, venous phase, portal venous phase, and delayed phase); (c) diameter statistics. The Lesions are categorized into large and small based on a diameter threshold of 2 cm; (d) diameter distribution; (e)-(h) an example of MSWAL: (e) axial plane; (f) coronal plane; (g) 3D show; (h) sagittal plane.
  • Figure 2: The left architecture is Inception nnU-Net, whose Bottleneck Block is Encoder Block apart from Inception Downsampling. (a) Mini Inception, a component of Inception nnU-Net (b) Inception Downsampling, another module of Inception nnU-Net. Both Mini Inception and Inception Downsampling have two branches, called branch left and branch right.
  • Figure 3: Visualization of the segmentation results from different methods. The yellow arrows point to the shortcomings of the six cutting-edge models in terms of edge segmentation, or to the regions of false positives and false negatives.