DeepValve: The first automatic detection pipeline for the mitral valve in Cardiac Magnetic Resonance imaging

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Accurate mitral valve (MV) assessment is crucial for diagnosing valvular disease and preventing complications.
• Cardiac magnetic resonance (CMR) provides detailed MV structure and function imaging, surpassing other modalities.
• Automated MV leaflet detection in CMR could significantly enhance diagnostic speed and accuracy.

Purpose of the Study:

• To introduce DeepValve, the first deep learning (DL) pipeline for automated MV detection in CMR.
• To evaluate three DL models (UNET-REG, UNET-SEG, DSNT-REG) for MV leaflet detection and localization.
• To propose novel metrics for assessing MV detection quality in CMR.

Main Methods:

• Developed and tested the DeepValve DL pipeline on 120 clinical CMR images from patients with MV disease.
• Implemented and compared a keypoint-regression model (UNET-REG), a segmentation model (UNET-SEG), and a hybrid keypoint model (DSNT-REG).
• Utilized Procrustes-based and customized Dice-based metrics for evaluating model performance.

Main Results:

• The DSNT-REG model demonstrated superior regression performance, accurately identifying MV landmark locations.
• The UNET-SEG model achieved satisfactory Dice scores, accurately predicting MV location and topology.
• Both models showed effectiveness in different aspects of MV assessment.

Conclusions:

• DeepValve represents a significant advancement in automated MV assessment using DL in CMR.
• The developed models offer potential for rapid and precise clinical evaluation of MV disease.
• This work paves the way for improved diagnostic workflows and patient care in valvular heart disease.