Author: Jiali Gong, Yi Guo, Chi Han, Wanli Huo, Hongdong Liu, Zhao Peng, Yaping Qi, Zhaojuan Zhang 👨🔬
Affiliation: Department of Radiotherapy, cancer center, The First Affiliated Hospital of Fujian Medical University, Department of Oncology, Xiangya Hospital, Central South University, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, China Jiliang University, Division of lonizing Radiation Metrology, National Institute of Metrology, Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, College of Information Engineering, China Jiliang University, the Zhejiang-New Zealand Joint Vision-Based Intelligent Metrology Laboratory, College of Information Engineering, China Jiliang University 🌍
Purpose: To address overfitting from limited training data in multi-organ segmentation, an efficient transfer learning framework is proposed. It reduces reliance on training samples, enabling a single organ model to perform multi-organ segmentation with minimal training, improving efficiency.
Methods:
A multi-organ segmentation framework based on cross-domain transfer learning is proposed. A model pre-trained on ImageNet extracts generic features, which are adapted to medical image segmentation. Data from a tumor dataset with a similar feature space is used for pre-training, and model parameters are transferred to multiple organ segmentation tasks, fine-tuned with minimal labeled data. A multi-scale feature fusion hybrid embedding model is designed to enhance segmentation accuracy and spatial consistency. This framework significantly improves the accuracy and robustness of multi-organ segmentation through effective cross-domain transfer.
Results:
Experimental results show that the dual cross-domain transfer strategy significantly improves segmentation performance on clinical datasets, with the average Dice coefficient for the bladder, rectum, and small intestine exceeding 90%, and the mean IoU also surpassing 90%. The proposed hybrid embedding model fully utilizes cross-domain transfer advantages, outperforming other models across all metrics and demonstrating excellent segmentation accuracy and generalization.
Conclusion: This work shows that the cross-domain transfer learning-based multi-organ segmentation method reduces reliance on training samples, enabling a single organ model to perform multi-organ segmentation with minimal training. It improves training efficiency, segmentation performance, and robustness, addressing challenges like data scarcity and overfitting. By combining model and data cross-domain transfer, the method leverages existing domain knowledge, enhancing accuracy and generalization. The hybrid embedding model excels in various medical image segmentation tasks, proving its clinical potential. This innovative and feasible approach offers a reliable solution with practical application value in medical image segmentation.