Fine-grained calibrated double-attention convolutional network for left ventricular segmentation.

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Deep Learning

Background:

• Accurate left ventricular (LV) segmentation in cardiac MRI is crucial for early heart disease diagnosis and treatment.
• Convolutional neural networks can lose critical LV information during segmentation, impacting clinical parameter accuracy.
• Precise segmentation of the ventricular contour remains a significant challenge.

Purpose of the Study:

• To introduce a novel deep learning model, the fine-grained calibrated double-attention network (FCDA-Net), for precise LV endocardium and epicardium segmentation.
• To enhance the similarity between neural network outputs and target segmentation regions in cardiac MRI.
• To improve the accuracy of clinical parameter calculations derived from LV segmentation.

Main Methods:

• FCDA-Net utilizes a U-net architecture as its backbone.
• The network incorporates a double grouped-attention module, comprising fine calibration spatial attention (fcSAM) and fine calibration channel attention (fcCAM) modules.
• This attention mechanism refines spatial and channel-wise feature maps for enhanced segmentation calibration.

Main Results:

• The FCDA-Net model demonstrated superior LV segmentation performance compared to existing advanced methods on the MICCAI 2009 challenge dataset.
• Ablation experiments confirmed the effectiveness of individual grouped-attention modules within the FCDA-Net architecture.
• The proposed method achieved better segmentation results, indicating improved accuracy.

Conclusions:

• FCDA-Net offers a significant advancement in LV segmentation from cardiac MRI.
• Improved segmentation accuracy facilitates more reliable quantitative analysis of essential clinical parameters.
• This technology holds potential for enhancing image-guided clinical surgery.