Novel Domain Knowledge-Encoding Algorithm Enables Label-Efficient Deep Learning for Cardiac CT Segmentation to Guide...

Area of Science:

• Medical Imaging and Machine Learning
• Computational Anatomy and Cardiac Segmentation

Background:

• Accurate segmentation of computed tomography (CT) scans is vital for personalized cardiac ablation procedures targeting cardiac arrhythmias.
• Current machine learning (ML) approaches for automated segmentation are limited by the requirement for large, meticulously labeled training datasets, which are difficult and time-consuming to acquire.
• This limitation hinders the widespread clinical adoption of ML-based segmentation tools.

Purpose of the Study:

• To develop and validate a novel automated labeling approach, the domain knowledge-encoding (DOKEN) algorithm, to enable high-performance ML segmentation from small training sets.
• To reduce the dependency on extensive labeled data by encoding cardiac geometry and anatomical knowledge.
• To assess the efficacy of the DOKEN algorithm in segmenting cardiac structures for atrial fibrillation (AF) ablation procedures.

Main Methods:

• The DOKEN algorithm was developed to parse left atrial (LA) structures and extract anatomical knowledge from publicly available digital LA models.
• This extracted knowledge was used to automatically label a small training dataset.
• A nnU-Net deep neural network (DNN) model was trained using the DOKEN-labeled data and subsequently tested on a larger hold-out dataset (N=100 patients) from an AF ablation study.

Main Results:

• The DOKEN-enhanced nnU-Net model achieved high segmentation performance with a training-to-test ratio of 1:5, demonstrating a Dice score of 96.7%.
• Boundary segmentation accuracy was excellent, with a median surface distance error of 1.51 mm and a mean centroid-boundary distance of 1.16 mm, comparable to expert performance (r = 0.99).
• In digital heart models, DOKEN achieved a mean centroid-boundary distance difference of -0.27 mm (r = 0.99), validating its precision in anatomical segmentation.

Conclusions:

• The DOKEN algorithm effectively performs segmentation of multiple left atrial substructures, significantly reducing the need for large training datasets in ML.
• This novel integration of domain knowledge with ML offers a promising solution to overcome data limitations in medical image analysis.
• The approach holds potential for application in AF ablation, with future extensions to other imaging modalities, 3D printing, and data science fields.