Effects of data-centric multi-task learning with larger patch sizes on pulmonary nodule segmentation performance(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第10期
- Page:
- 1306-1320
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Effects of data-centric multi-task learning with larger patch sizes on pulmonary nodule segmentation performance
- Author(s):
- LIU Jian1; ZHANG Zheng2; NIU Bing3; KANG Shuai3; REN Juan3; WANG Lei4; XU Kai1
- 1. School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200241, China 2. Department of Radiology, Shanghai Pulmonary Hospital, Tongji University, Shanghai 200433, China 3. Shanghai Simple Touch Technology Co., Ltd., Shanghai 201600, China 4. Department of Oncology, Shanghai Pulmonary Hospital, Tongji University, Shanghai 200433, China
- Keywords:
- Keywords: pulmonary nodule multi-task learning deep learning segmentation performance
- PACS:
- R318；R445.3；TP181
- DOI:
- DOI:10.3969/j.issn.1005-202X.2025.10.007
- Abstract:
- Abstract: Given the lack of annotations for key lung organs and tissues in existing public datasets, this study collected 863 cases of chest CT scan images and constructed the first comprehensive dataset containing annotations of pulmonary vessels, airways, and nodules using a semi-automated method that combines computer vision algorithms with manual corrections by radiologists. On this basis, a lung nodule segmentation model based on multi-task learning is proposed. By incorporating annotations of pulmonary vessels (pulmonary arteries and veins) and the trachea to enhance models ability to learn lung features, the proposed model reduces the false discovery rate in lung nodule detection, and improves generalization ability. Additionally, the use of larger image patches further optimizes model performance. The trained VAAN_128 model achieves the best performance, with a Dice coefficient of 0.694 and a false discovery rate of 0.210 for lung nodule segmentation. Moreover, it simultaneously provides accurate segmentation results of pulmonary vessels and the trachea, assisting in the formulation of more precise diagnosis and treatment plans. Based on the VAAN_128 model, a software system for navigation and localization in biopsy procedures is developed. In clinical practice, this system can assist physicians in accurately locating lung nodules, distinguishing critical tissues, and improving preoperative planning efficiency. This provides precise and efficient technical support for early diagnosis and disease monitoring of lung diseases, and is of great significance for path planning in clinical navigation system and future lung imaging research.
Last Update: 2025-10-29