IEU-Net: Sequence learning of internal texture and external morphology for sonographic kidney segmentation.
Authors
Affiliations (5)
Affiliations (5)
- Institute of Artificial Intelligence Innovation, Industry Academia Innovation School, National Yang Ming Chiao Tung University, Hsinchu, Taiwan. Electronic address: [email protected].
- Department of Mathematical Sciences, National Chengchi University, Taipei, Taiwan.
- Division of Nephrology, Department of Internal Medicine, Kaohsiung Medical University Hospital, Kaohsiung Medical University, Kaohsiung, Taiwan.
- Department of Mathematical Sciences, National Chengchi University, Taipei, Taiwan. Electronic address: [email protected].
- Division of Nephrology, Department of Internal Medicine, Kaohsiung Medical University Hospital, Kaohsiung Medical University, Kaohsiung, Taiwan; Biomedical Artificial Intelligence Academy, Faculty of Medicine, College of Medicine, Kaohsiung Medical University, Kaohsiung, Taiwan; Center for Big Data Research, Kaohsiung Medical University, Kaohsiung, Taiwan. Electronic address: [email protected].
Abstract
The global prevalence rate of chronic kidney disease (CKD) is on the rise, posing a significant public health concern. Renal ultrasound is an important assessment tool for clinicians. Using automatic ultrasound kidney segmentation would be helpful for kidney evaluation. However, challenges arise due to the limited data size and variabilities in quality across population datasets for kidneys. Addressing the issues, this study developed IEU-Net, a sequence learning architecture to integrate multi-center datasets and comprehensive features. The datasets were collected from different countries, including China, the United States, Canada, and Taiwan. IEU-Net was composed of an original nnU-Net and a modified nnU-Net to learn internal texture and external morphology features. IEU-Net achieved the best dice similarity coefficient (DSC) of 0.8832, compared to DSC values of other networks. Based on the foundation of nnU-Net, IEU-Net achieved a 6.95% accuracy improvement with a sequence learning architecture. When the backbone of this architecture was replaced with other networks, their improvements ranged from 3.64% to 20.13%. These findings contribute to the growing body of evidence supporting the practical application of deep learning-based segmentation in kidney ultrasound, underscoring its potential utility in clinical settings. The code is available at https://github.com/cvrlab308/IEU_NET.