Analyzing right ventricular attributes with machine learning: A deep learning solution trained on model-annotated data

Abstract

An accurate analysis of echocardiograms is key to diagnosing heart conditions in modern medicine. This analysis is today performed by physicians and requires expertise that might only come after years of training in the area. In this master thesis we explore the possibility of using machine learning models, trained on large amounts of echocardigrams, that when trained can predict the mobility and size of the right ventricle. A successful model could be useful in assisting both inexperienced physicians during training or as a tool to speed up analysis for the already trained experts. To train a supervised model, large amounts of annotated data is required but annotating the echocardiograms requires high expertise and is time consuming. As a possible solution to this problem, a separate text classifier trained on transcript data were used to annotate the echocardiograms that were linked to the respective examination. The transcript data consisted of a physicians analysis on all examination data that belonged to a patient, including the echocardiograms. Several different architectures for the text classifier were trained and evaluated and the best performing model achived 92% accuracy on classifying the mobility and 95% accuracy at classifying the size of the right ventricle using transcript data. The trained text models were then used to annotate the echocardiograms in the dataset and the resulting data were then was used to train a set of image classifiers. The echocardiograms are 3-dimensional data and our results showed that models using a 3-dimensional representation also performed the best on the two classification tasks. The best models used a combination of human-annotated and model-annotated data and achieved 82% and 83% accuracy on mobility and size respectively. This result can be compared to the interobserver agreements between our transcript analysis and two experts annotating echocardiograms from the test set which were 82% for mobility and 72% for size. The study showed that the use of a machine learning model as tool for physicians is a feasible and an interesting prospect. As a continuation of the work done here, the first step would be to increase the size of the models and training data. With the help of the automatic annotation, additional data is relatively easy to process. Other ways forward would be to include additional parameters from the examinations such as heart rate and to analyze the right ventricle attributes in relation to each other.