Impact of fine-grained learning rate configuration on the performance of medical image segmentation models.
Authors
Affiliations (4)
Affiliations (4)
- School of Computer Science and Technology, Taiyuan University of Science and Technology, No. 66 Rangliu Road, Wanbailin District, Taiyuan City, Shanxi Province, China, Taiyuan, 030024, CHINA.
- Northeastern University, Building B, Life Building, 195 Innovation Road, Hunnan New District, Shenyang, Liaoning, China, Shenyang, Liaoning, 110016, CHINA.
- Taiyuan University of Science and Technology, 66wa Liu Road, Wanbolin District, Taiyuan City, Shanxi Province, China, Taiyuan, Shanxi , 030024, CHINA.
- The College of Computer Science and Technology, Taiyuan University of Science and Technology, 66wa Liu Road, Wanbolin District, Taiyuan City, Shanxi Province, China, Taiyuan, Shanxi , 030024, CHINA.
Abstract
Research on deep learning for medical image segmentation has shifted from single-modality networks to multimodal data fusion. Updating the parameters of such deep learning models is crucial for accurate segmentation predictions. Although existing optimizers can perform global parameter updates, the fine-grained initialization of learning rates across different network hierarchies and its influence on segmentation performance has not been sufficiently explored. To address this, we conducted a series of experiments showing that the initialization of a differentiated learning rate across network layers directly affected the performance of medical image segmentation models. To determine the optimal initial learning rate for each network level, we summarized a general statistical relationship between early-stage training results and the model's final optimal performance. In this paper, we proposed a fine-grained learning rate configuration algorithm. To verify the effectiveness of the proposed algorithm, we evaluated 10 segmentation models on three benchmark datasets: the colon polyp segmentation dataset CVC-ClinicDB, the gastrointestinal polyp dataset Kvasir-SEG, and the breast tumor segmentation dataset BUSI. The models that achieved the most significant improvement in mIoU on these three datasets were H-vmunet, MSRUNet, and H-vmunet, with increases of 3.87%, 4.67%, and 6.22%, respectively. Additionally, we validated the generalization and transferability of the proposed algorithm using a thyroid nodule segmentation dataset and a skin lesion segmentation dataset. Finally, a series of analyses, including segmentation result analysis, feature map visualization, training process analysis, computational overhead analysis, and clinical relevance analysis, confirmed the effectiveness of the proposed method. The core code is publicly available at https://github.com/Lambda-Wave/PaperCoreCode.