Related Experiment Video
Updated: Oct 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
TSU-net: Two-stage multi-scale cascade and multi-field fusion U-net for right ventricular segmentation
Xiuquan Du1, Xiaofei Xu2, Heng Liu3
1Key Laboratory of Intelligent Computing and Signal Processing, Ministry of Education, Anhui University, Hefei, Anhui, China; School of Computer Science and Technology, Anhui University, Hefei, Anhui, China.
Summary
TSU-net improves right ventricle segmentation in cardiac MRI using novel Dilated-Convolution and Multi-Layer-Pool blocks. This method enhances accuracy for cardiac function analysis and disease diagnosis.
Area of Science:
- Medical Imaging
- Cardiovascular Imaging
- Artificial Intelligence in Medicine
Background:
- Accurate segmentation of the right ventricle (RV) in cardiac magnetic resonance images (MRI) is crucial for cardiac function analysis and disease diagnosis.
- Challenges in RV segmentation include variations in object size and ill-defined borders.
- Existing methods struggle to effectively capture multi-scale features and handle boundary ambiguities.
Purpose of the Study:
- To introduce TSU-net, a novel deep learning network for accurate right ventricle segmentation in cardiac MRI.
- To address the limitations of existing methods in handling varying object sizes and ill-defined borders.
- To improve the precision of cardiac function analysis and disease diagnosis through enhanced RV segmentation.
Main Methods:
- Developed TSU-net, incorporating Dilated-Convolution Blocks (DB) and Multi-Layer-Pool Blocks (MB).
- DB extracts and aggregates multi-scale features; MB detects objects of different sizes and fills boundary features using multi-field-of-views.
- Integrated DB and MB into both encoding and decoding layers of a two-stage U-net structure to enhance multi-scale information processing and boundary feature detection.
Main Results:
- TSU-net achieved high performance on the RVSC dataset, with an average Dice coefficient of 0.86 for the endocardium and 0.90 for the epicardium.
- The proposed network outperformed existing segmentation models.
- The network demonstrated effectiveness in detecting targets of different sizes and filling boundary features.
Conclusions:
- TSU-net offers a significant advancement in right ventricle segmentation accuracy from cardiac MRI.
- The method provides an effective tool for assisting clinicians in disease diagnosis and advancing medical imaging analysis.
- The network's architecture successfully addresses challenges related to multi-scale features and boundary definition in RV segmentation.
