Penerapan Deep Learning untuk Klasifikasi Penyakit Katup Jantung pada Video Ekokardiografi Doppler

Abstract

Valvular heart diseases, such as mitral regurgitation (the backflow of blood from the ventricle into the left atrium) and mitral stenosis (narrowing of the valve that impedes blood flow), can disrupt the hemodynamic function of the heart. Doppler echocardiography serves as the primary diagnostic modality for detecting these conditions; however, its interpretation remains highly dependent on the expertise of cardiologists, whose availability is limited in Indonesia. This study develops an automated classification system based on Doppler echocardiography videos to distinguish between two types of mitral valve disorders (regurgitation and stenosis) and normal conditions. Video-based classification poses unique challenges due to the need for simultaneous processing of spatial and temporal information, which is inherently more complex than static image analysis. Videos are converted into arrays of frames, and latent features are extracted using a pretrained autoencoder-based ConvNeXt-Tiny encoder enhanced with a CBAM module. Four Top-K frame selection methods are compared: Temporal Change (delta-norm), a combination of Feature Strength and Temporal Change, a combination of Feature Strength and Uniqueness, and K-Means Clustering. The model was evaluated using stratified 5-fold cross-validation on the training portion of the dataset, followed by a final assessment on a separate hold-out test set. The classification model architecture consists of the autoencoder-based encoder, BiLSTM, dense attention, and fully connected layers. The Temporal Change method achieves the best performance with an average accuracy of 0.80±0.05 and an ROC AUC of 0.91±0.03. Final testing results yield an accuracy of 0.85, precision of 0.87, recall of 0.85, F1-score of 0.85, ROC AUC of 0.90, MCC of 0.77, Cohen’s kappa of 0.77, and specificity values of 0.92 (mitral regurgitation), 1.00 (mitral stenosis), and 0.84 (normal). The findings indicate that the proposed deep learning method performs effectively and consistently in identifying mitral valve disorders, suggesting its viability for future use as part of a clinical decision support system. Furthermore, a prototype web application was developed to serve as a proof-of-concept, enabling real-time processing of echocardiographic video data.