Automatic delineation of cardiac substructures using a region-based fully convolutional network.

Area of Science:

• Medical imaging
• Artificial intelligence
• Radiotherapy

Background:

• Radiation dose to cardiac substructures is linked to post-treatment toxicity.
• Substructure dose is more predictive of toxicity than whole-heart dose.
• Accurate cardiac substructure segmentation is crucial for dosimetric studies.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for automatic cardiac substructure contouring on CT scans.
• To improve the understanding of radiation dose-toxicity relationships in cardiac substructures.
• To provide a tool for retrospective and prospective dosimetric studies.

Main Methods:

• A mask-scoring regional convolutional neural network (RCNN) was employed, featuring a backbone network, RPN, RCNN head, mask head, and mask-scoring head.
• The network was trained on 55 patient CT datasets and evaluated using threefold cross-validation and a holdout cohort.
• Performance was compared against a 3D UNet, assessing metrics like Dice score coefficients (DSCs) and mean surface distances (MSDs).

Main Results:

• The deep learning method achieved high DSCs for whole heart (0.96), chambers (0.94), and great vessels (0.93), outperforming the 3D UNet.
• MSDs were <2 mm for most substructures, indicating precise segmentation.
• The algorithm demonstrated statistically significant improvements with contrast imaging for whole heart and chambers, and rapid segmentation (<5 s per dataset).

Conclusions:

• A deep learning network effectively performs automatic cardiac substructure delineation using CT data.
• This automated approach can significantly aid in investigating cardiac substructure radiation dose and treatment toxicities.
• The tool facilitates more accurate and efficient dosimetric analyses for radiotherapy planning and research.