Cardiac segmentation on CT Images through shape-aware contour attentions

Area of Science:

• Medical Image Analysis
• Deep Learning
• Cardiovascular Imaging

Background:

• Accurate cardiac segmentation in computed tomography (CT) images is crucial for early cardiovascular disease diagnosis.
• Deep learning models have shown promise in medical image segmentation but struggle with proximate cardiac substructures due to indiscernible boundaries.
• Existing methods face challenges in focusing segmentation networks on the boundaries between cardiac chambers and vessels.

Purpose of the Study:

• To develop a novel deep learning model for improved cardiac segmentation accuracy in CT images.
• To enhance the model's ability to exploit shape and boundary-aware features for better differentiation of proximate cardiac substructures.
• To outperform conventional contour-based attention methods in segmenting complex cardiac anatomy.

Main Methods:

• Introduction of a novel shape-aware attention module that utilizes distance regression.
• The proposed module guides the model to focus on the edges and boundaries between cardiac substructures.
• Comparison with conventional contour-based attention methods to evaluate performance.

Main Results:

• The proposed network achieved a 4.97% improvement in the Dice similarity coefficient score compared to state-of-the-art networks.
• Experiments were conducted on the Multi-Modality Whole Heart Segmentation dataset (20 images for training/validation, 40 for testing).
• The model demonstrated more accurate segmentation results on CT cardiac images.

Conclusions:

• The shape-aware contour attention mechanism effectively utilizes distance transformation and boundary features to improve attention maps.
• The method strengthens responses in boundary areas, leading to more accurate segmentation of cardiac substructures.
• The proposed approach significantly reduces false-positive responses, enhancing the overall accuracy of cardiac segmentation.