Table 2.

Non-commercial AI-driven algorithms for valvular detection in echocardiography

Authors (year)	Data	Training data population	Outcome of interest	Algorithm used	Findings
Vafaeezadeh et al (2021)35	2044 TTE studies: 1597 had normal valves and 447 had prosthetic valves.	Patients with normal mitral valve and mitral valve prosthesis: both mechanical and biological.	Identification of prosthetic mitral valve from echo images.	13 pretrained models with CNN architecture and fine-tuned via transfer learning.	All the models worked with incredible accuracy (>98%), but the EfficientNetB3 had the best AUC (99%) for the A4C and EfficientNetB4 had the best AUC (99%) for PLAX. However, these models were computationally more expensive for a small gain in AUC, so the authors concluded that the best model for this task is EfficientNetB2.
Corinzia et al (2020)50	Training: 39 2D TTE. Test: 46 2D echos from EchoNet-Dynamic public echo data set.	Patients who were undergoing mitraclip: all patients had moderate to severe or severe MR.	Fully automated delineation of mitral valve annulus and both MV leaflets.	NN-MitralSeg, unsupervised MV segmentation algorithm based on neural collaborative filtering.	This model outperforms state-of-the-art unsupervised and supervised methods (NeuMF MF Dice coefficient of 0.482, with benchmark performance of 0.447), with best performance on low-quality videos or videos with sparse annotation.
Andreassen et al (2020)29	111 multiframe recordings from 3D TEE echocardiograms.	4D echocardiographic images of the mitral valve.	Fully automated method for mitral annulus segmentation on 3D echocardiography.	CNN, specifically a U-Net architecture.	With no manual input, this methodology gave comparable results with those that required manual input (relative error of 6.1%±4.5% for perimeter measurements and 11.94%±10% for area measurement).
Costa et al (2019)28	Training: 21 2D TTE echo videos in PLAX, 22 videos in A4C. Test: 6 videos in PLAX and A4C.	PLAX and A4C views from echos.	Automatic segmentation of mitral valve leaflets.	CNN, specifically a U-Net architecture.	This model is the first of its kind to perform segmentation of valve leaflets. For AMVL, the median Dice coefficient in PLAX was 0.742 and 0.795 in A4C. For PMVL, the median Dice coefficient in PLAX was 0.60 and 0.69 in A4C. Cardiologists were then asked to score the segmentation quality on a scale from 0 to 2, with pooled score of 0.781, suggesting reasonable quality segmentation.

A4C, apical 4-chamber view; AI, artificial intelligence; AML, anterior mitral valve leaflet; AUC, area under the receiver operator curve; CNN, convolutional neural network; 2D, two-dimensional; 3D, three-dimensional; 4D, four-dimensional; MR, mitral regurgitation; MV, mitral valve; PLAX, parasternal long-axis view; TEE, transoesophageal echocardiogram; TTE, transthoracic echocardiogram.