Related Experiment Video
Updated: Nov 2, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Cardiac MRI segmentation with focal loss constrained deep residual networks
Chuchen Li1, Mingqiang Chen1, Jinglin Zhang2
1College of Optical Science and Engineering, Zhejiang University, People's Republic of China.
Insights
This study introduces a novel FR-Net model to improve cardiac magnetic resonance imaging segmentation by addressing challenges with small, hard-to-segment heart structures. The method enhances accuracy by re-weighting gradients and combining focal and dice losses for better optimization.
Area of Science:
- Medical Imaging
- Computer Vision
- Biomedical Engineering
Background:
- Accurate delineation of cardiac structures in cardiac magnetic resonance imaging (CMRI) is essential for diagnosis and treatment.
- Automatic segmentation of heart structures, particularly ventricles in apex slices, is challenging due to small target sizes and imbalanced datasets.
- Standard training methods struggle with hard-to-segment samples (e.g., endocardium occupying ~4% of image area), leading to optimization bias towards easy background samples.
Purpose of the Study:
- To develop an improved method for automatic segmentation of cardiac structures in CMRI, specifically addressing the challenge of small and difficult-to-segment regions.
- To enhance the accuracy and robustness of cardiac segmentation models by mitigating the impact of imbalanced data and hard training samples.
- To introduce a novel deep learning architecture, the focal loss constrained residual network (FR-Net), for precise CMRI segmentation.
Main Methods:
- Proposed a focal loss constrained residual network (FR-Net) incorporating a pixel-wise re-weighting strategy to balance gradients from easy and hard samples.
- Implemented an alternative training approach that alternately applies focal loss (pixel-wise) and dice loss (region-wise) for comprehensive model optimization.
- Evaluated the FR-Net model on multiple datasets, including Sunnybrook, CMRI, right ventricle, and ACDC datasets.
Main Results:
- The FR-Net model demonstrated effectiveness in improving the segmentation of challenging cardiac structures, particularly in images with small target regions.
- The combination of pixel-wise focal loss and region-wise dice loss led to significant improvements in segmentation accuracy.
- Experimental results across various datasets validated the proposed method's performance.
Conclusions:
- The proposed FR-Net with its innovative training strategy effectively addresses the limitations of existing methods in CMRI segmentation.
- The approach enhances the ability to accurately segment small and difficult cardiac structures, benefiting clinical applications.
- This work offers a promising solution for improving automated cardiac segmentation in medical imaging.
Abstract:
Delineating anatomical structures for cardiac magnetic resonance imaging (CMRI) is crucial for various medical applications such as medical diagnoses, treatment, and pathological studies. CMRI segmentation, which aims to automatically and accurately segment the heart structures, is highly beneficial for cardiologists. However, it is non-trivial to perfectly segment the ventricles, especially for the heart apex slices, considering their small sizes compared to the input images. For example, the endocardium in the Sunnybrook dataset only occupies 4% of the entire image by average. During the training process, these target pixels, or other hard samples, are buried by the massive backgrounds that make the model mostly receive optimization signals from easy samples. In this paper, we propose a focal loss constrained residual network (FR-Net) to tackle the problem. In order to mitigate the fact that the gradients of the hard samples can be easily overwhelmed by the easy samples, we use a pixel-wise re-weighting strategy to balance the gradients. Furthermore, considering focal loss constraints for each pixel independently, we propose an alternative training fashion that trains the model with focal loss and dice loss alternatively. The segmentation model can not only benefit from the pixel-wise focal loss but also from the region-wise dice loss to comprehensively optimize the model. We conducted thorough experiments on the Sunnybrook dataset, CMRI dataset, right ventricle dataset, and ACDC dataset to verify the effectiveness of the proposed method.
