Fully automated cardiac MRI segmentation using dilated residual network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Cardiac ventricle segmentation from CMRI is crucial for assessing cardiovascular diseases.
• Deep learning models like UNet show promise but struggle with feature loss in the bottleneck layer.
• Improving feature representation at the UNet bottleneck is key to enhancing segmentation performance.

Purpose of the Study:

• To enhance the performance of cardiac ventricle segmentation by addressing feature degradation at the UNet bottleneck.
• To develop a fully automatic pipeline for segmenting right ventricle (RV), myocardium (MYO), and left ventricle (LV) from CMRI.
• To improve spatial and temporal information capture and localization accuracy in cardiac segmentation.

Main Methods:

• A novel dilated residual network (DRN) was integrated into the UNet bottleneck to capture features at full resolution.
• A data-augmentation technique was employed to mitigate overfitting and class imbalance issues.
• Pixel-wise addition of outputs from expanding paths was used to refine the training response.

Main Results:

• The proposed DRN method achieved high Dice Similarity Coefficient (DSC) scores: 0.924 ± 0.03 (RV), 0.907 ± 0.01 (MYO), and 0.949 ± 0.05 (LV).
• Excellent Hausdorff Distance (HD) scores were obtained: 10.09 ± 0.01 mm (RV), 7.25 ± 0.05 mm (MYO), and 6.86 ± 0.02 mm (LV).
• The model demonstrated superior performance compared to state-of-the-art methods, with overall DSC and HD improvements of 1.0% and 1.5%, respectively.

Conclusions:

• The developed dilated residual UNet (DRN) effectively segments cardiac ventricles from short-axis CMRI.
• The method excels at restoring spatial-temporal information and preserving image features without degradation.
• The highly accurate and rapid segmentation (0.28s/image) surpasses existing methods for ventricular region analysis.