Deep Learning for Automated Measurement of Total Cardiac Volume for Heart Transplantation Size Matching.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Total Cardiac Volume (TCV) is crucial for pediatric heart transplant graft selection.
• Manual TCV segmentation from CT scans is time-consuming and requires specialized expertise.
• Automating TCV calculation can enhance graft utilization and streamline the transplant process.

Purpose of the Study:

• To develop and validate a Deep Learning (DL) model for accurate TCV calculation from CT images.
• To assess the feasibility of using a 3-dimensional Convolutional Neural Network (3D-CNN) for automated TCV measurement.
• To enable faster and more widespread use of TCV in pediatric heart transplantation.

Main Methods:

• A custom 3D-CNN model integrating DenseNet and ResNet architectures was developed.
• The model was trained on CT scans from 270 subjects and validated on 44 subjects (aged 0-30 years).
• Ground truth TCV was established through manual segmentation for training and testing.

Main Results:

• The DL model achieved a high average Dice similarity coefficient of 0.94 ± 0.03.
• Mean absolute percent error for TCV estimation was 5.5%, with no significant age, weight, or height correlation.
• The model demonstrated higher accuracy for normal hearts compared to those listed for transplant.

Conclusions:

• A DL-based 3D-CNN model can accurately and automatically measure TCV from CT images.
• This automated approach has the potential to significantly improve pediatric heart transplant graft matching.
• Further multicenter studies are needed to generalize findings and enhance accuracy with diverse data.