This study presents the development and validation of a fully automated deep learning pipeline for detecting clinically significant mitral regurgitation (MR) from transthoracic echocardiography (TTE) studies. The pipeline was trained on 58,614 TTE studies from Cedars-Sinai Medical Center (CSMC) and tested on 1,800 internal studies and 915 external studies from Stanford Healthcare (SHC). The model successfully identified apical-4-chamber (A4C) videos with color Doppler across the mitral valve and assessed MR severity with high accuracy. In the internal test set, the view classifier achieved an AUC of 0.998 and correctly identified 3,452 of 3,539 MR color Doppler videos. In the external test set, the view classifier correctly identified 1,051 of 1,055 MR color Doppler videos. For MR severity assessment, the model detected moderate-or-severe MR with an AUC of 0.916 in CSMC and 0.951 in SHC, and severe MR with an AUC of 0.934 in CSMC and 0.969 in SHC. The model demonstrated strong performance in distinguishing MR severity and identifying clinically significant MR. The model's performance was consistent across institutions, with high sensitivity and specificity. The model's interpretability was evaluated using saliency maps, which highlighted clinically relevant imaging features of MR. The study concludes that the developed deep learning pipeline can automate the detection and assessment of MR, potentially aiding in screening and surveillance. The model's high AUC, NPV, and generalizability across sites suggest its potential for use in primary care settings and low-resource environments. The study also highlights the importance of integrating AI into echocardiography for improving diagnostic accuracy and efficiency.This study presents the development and validation of a fully automated deep learning pipeline for detecting clinically significant mitral regurgitation (MR) from transthoracic echocardiography (TTE) studies. The pipeline was trained on 58,614 TTE studies from Cedars-Sinai Medical Center (CSMC) and tested on 1,800 internal studies and 915 external studies from Stanford Healthcare (SHC). The model successfully identified apical-4-chamber (A4C) videos with color Doppler across the mitral valve and assessed MR severity with high accuracy. In the internal test set, the view classifier achieved an AUC of 0.998 and correctly identified 3,452 of 3,539 MR color Doppler videos. In the external test set, the view classifier correctly identified 1,051 of 1,055 MR color Doppler videos. For MR severity assessment, the model detected moderate-or-severe MR with an AUC of 0.916 in CSMC and 0.951 in SHC, and severe MR with an AUC of 0.934 in CSMC and 0.969 in SHC. The model demonstrated strong performance in distinguishing MR severity and identifying clinically significant MR. The model's performance was consistent across institutions, with high sensitivity and specificity. The model's interpretability was evaluated using saliency maps, which highlighted clinically relevant imaging features of MR. The study concludes that the developed deep learning pipeline can automate the detection and assessment of MR, potentially aiding in screening and surveillance. The model's high AUC, NPV, and generalizability across sites suggest its potential for use in primary care settings and low-resource environments. The study also highlights the importance of integrating AI into echocardiography for improving diagnostic accuracy and efficiency.