Echocardiographic dataset creation and left ventricular hypertrophy detection using a weakly supervised residual neural network

Abstract

Left ventricular hypertrophy (LVH) is a cardiac structural change characterized by an increase in the ventricular wall mass which can lead to heart failure. The aim of this work is to use deep learning to automatically detect left ventricular hypertrophy from echocardiograms.

We collected a dataset of about 10,000 images and built a single-image ResNet50-based classifier to detect LVH. Furthermore, we applied Grad-CAM analysis to obtain a visual validation of the model.

The network achieved an AUC of 0.99, an accuracy of 0.94, and an F1-score of 0.94. Grad‑CAM analisis confirmed that the model focused on regions relevant for the LVH diagnosis. In conclusion, our network has the ability to automatically detect LVH and also to localize key cardiac structures with only image-level labels as supervision.

Objective

Create a data structure suitable for machine learning tasks and detect left ventricular hypertrophy from echocardiography using deep learning.

Research methodology

We obtained a dataset of about 10,000 images which was split in: 80% for the training and validation sets used during a 3-fold cross validation training, and 20% for the testing set. We built and trained a ResNet50‑based classifier using Keras library of Python. We used Grad-CAM analysis to obtain a visual validation of the model.

Conclusions

The model achieved an accuracy of 0.94, an AUC of 0.99, and an F1-score of 0.94 on the test set. Grad-CAM analisis showed that the model focused on the posterior wall of the left ventricle which indeed is a relevant region for the diagnosis of LVH.

Future developments

Extend the dataset, introduce a segmentation step, select only certain frames from echocardiograms, detect other pathologies.