A fully convolutional neural network for the quantification of mitral regurgitation in echocardiography.

BACKGROUND

Mitral regurgitation (MR) is the most common form of valvular heart disease (VHD), and the accurate assessment of MR severity is critical for clinical management. However, the quantitative assessment of MR is intricate and time-consuming, posing challenges for physicians in ensuring the precision of the results. Thus, our objective was to create an automated and reproducible artificial intelligence (AI) system. This study aimed to assist physicians in grading MR severity using color Doppler echocardiograms through the implementation of a fully convolutional neural network (FCN).

METHODS

A retrospective cohort was established comprising 433 patients diagnosed with MR based on clinical criteria. Following screening, 269 patients met the inclusion criteria for the study. In total, 4,104 frames from apical 4-chamber view color Doppler flow images constituted the training and validation set, while 1,060 frames comprised the test set. Using the FCN, the MR flow convergence region was captured and segmented. The algorithm also estimated the parameter radius, which was employed to compute the effective regurgitant orifice area (EROA) and regurgitant volume (RV) based on the proximal isovelocity surface area. These measurements were subsequently graded following the 2017 American Society of Echocardiography (ASE) guidelines. The segmentation and grading performance of the model were assessed. Additionally, the diagnostic performance of the AI model was compared to that of ultrasound physicians with varying years of experience.

RESULTS

In groups I, II, III, and IV, the rates of correctly identifying the radius were 0.56, 0.83, 0.86, and 0.89, while the grading accuracy was 0.95, 0.89, 0.88, and 0.91, respectively. Regarding patients with MR of different etiologies, the grading accuracy for the functional MR and degenerative MR groups was 0.82 and 0.90, respectively. Using Carpentier classification of MR as the criterion, the accuracy for groups I, II, and IIIb was 0.80, 0.90, and 0.83, respectively.

CONCLUSIONS

The model showed commendable performance, streamlining the clinical diagnostic process and enhancing the precision and stability of quantitative MR assessment.