Multi-residual 2D network integrating spatial correlation for whole heart segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Whole heart segmentation (WHS) is crucial for cardiac analysis but faces challenges in accuracy, speed, and computational resources.
• Existing methods often struggle to meet the demands of practical clinical applications for cardiac CT imaging.

Purpose of the Study:

• To develop an accurate and efficient deep learning model for whole heart segmentation (WHS) on cardiac CT images.
• To address the need for fast inference speed and low GPU memory consumption in clinical settings.
• To improve the delineation of cardiac substructures through advanced feature extraction and attention mechanisms.

Main Methods:

• A novel 2D encoder-decoder network integrating spatial correlation for WHS on 3D cardiac CT images.
• Utilized a convolutional long short-term memory skip connection for spatial correlation feature extraction.
• Employed a multi-residual decoder with multi-scale and channel attention for refined feature analysis.

Main Results:

• Achieved high segmentation accuracy with Dice coefficient of 0.914 and Jaccard index of 0.843.
• Demonstrated fast inference time of 9.535 seconds and low GPU memory consumption of 1905 MB.
• Validated robustness and generalization across multiple datasets, including WHS and abdominal organ segmentation challenges.

Conclusions:

• The proposed multi-residual 2D network offers a highly accurate, efficient, and resource-conscious solution for WHS.
• The method shows strong potential for clinical deployment in cardiac imaging and adaptability to other organ segmentation tasks.
• Publicly available source code facilitates further research and development in medical image segmentation.