Related Experiment Video
Updated: Jan 5, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
An iterative multi-path fully convolutional neural network for automatic cardiac segmentation in cine MR images
Zongqing Ma1, Xi Wu2, Xin Wang3
1College of Computer Science, Sichuan University, Chengdu, Sichuan, 610065, China.
Medical Physics
|October 13, 2019
Summary
This study introduces an iterative multi-path fully convolutional network (IMFCN) for improved cardiac segmentation in cine MR images. The novel approach effectively utilizes spatial context, achieving state-of-the-art results for left ventricle (LV), right ventricle (RV), and myocardium (MYO) segmentation.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiology
Background:
- Accurate segmentation of cardiac structures like the left ventricle (LV), right ventricle (RV), and myocardium (MYO) is crucial for diagnosing and monitoring heart conditions using cine cardiac magnetic resonance (MR) imaging.
- Leveraging spatial context information can significantly enhance the performance of automated segmentation algorithms.
Purpose of the Study:
- To propose an iterative multi-path fully convolutional network (IMFCN) designed to effectively utilize spatial context for automatic cardiac segmentation in cine MR images.
- To improve the accuracy and consistency of LV, RV, and MYO segmentation by explicitly modeling interslice spatial correlations.
Main Methods:
- The IMFCN employs a multi-path late fusion strategy to process contextual inputs, including adjacent slices and previously predicted masks.
- An atrous spatial pyramid pooling (ASPP) module is integrated for effective fusion of high-level contextual features.
- Deep supervision (DS) and batch-wise class re-weighting mechanisms are utilized to optimize network training.
Main Results:
- The IMFCN demonstrated superior performance compared to methods lacking spatial context or using early fusion strategies on the MICCAI 2017 ACDC dataset.
- Achieved high Dice similarity coefficients (0.935 for LV, 0.920 for RV, 0.905 for MYO) and competitive Hausdorff distances on the ACDC test dataset.
- The model also showed comparable performance to state-of-the-art methods when retrained on the Sunnybrook dataset for LV segmentation, indicating broad applicability.
Conclusions:
- The developed IMFCN offers an effective, automatic, end-to-end solution for precise cardiac segmentation in cine MR images.
- The network's ability to leverage 2D spatial context leads to accurate and consistent segmentation outcomes, advancing cardiac image analysis.
