An Automated Heart Shunt Recognition Pipeline Using Deep Neural Networks.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Saline contrast transthoracic echocardiography (SC-TTE) is a key imaging modality for detecting heart shunts.
• Automated analysis of SC-TTE can improve diagnostic accuracy and accessibility for non-experts.
• Deep neural networks offer potential for developing sophisticated image analysis pipelines.

Purpose of the Study:

• To develop a fully automated and scalable analysis pipeline for distinguishing intracardiac and extracardiac heart shunts.
• To utilize a deep neural network-based framework for SC-TTE image analysis.
• To enable non-experts to accurately assess heart shunt lesions.

Main Methods:

• A three-step pipeline was developed: cardiac chamber segmentation (U-Net model), ultrasound microbubble localization, and disease classification.
• Multivariate time series data on microbubble counts were generated from segmented chambers and microbubble localization.
• A deep neural network classification model was trained on this data to differentiate shunt types.

Main Results:

• The framework achieved high accuracy in segmenting heart chambers (Dice coefficient = 0.92 ± 0.1) and localizing microbubbles.
• The disease classification model demonstrated high performance metrics for both intracardiac and extracardiac shunts.
• Specific performance metrics included accuracy (up to 0.902), sensitivity (up to 0.891), specificity (up to 0.966), F1 score (up to 0.836), kappa (up to 0.762), and AUC (up to 0.942).

Conclusions:

• The proposed deep neural network framework provides a fast, convenient, and accurate method for identifying intracardiac and extracardiac shunts.
• The automated pipeline aids in shunt recognition and generates valuable quantitative indices for clinical diagnosis.
• This technology has the potential to transform clinical practice by empowering non-experts in shunt assessment.