Artificial Intelligence applied to 2d echocardiography for the early detection of myocardial infarction

Abstract

Myocardial infarction (MI) occurs when blood flow to the heart muscle is blocked, usually as the result of underlying coronary artery disease (CAD). In 2019, ischemic CAD was responsible for 1 in 6 deaths globally, making it the principal cause of death worldwide. At present, diagnosis of heart disease, and specifically MI, requires particular biomarker and electrocardiography findings. However, since the findings of these methods are sometimes inconclusive and can result in misdiagnosis with other syndromes, cardiologists recurrently rely on imaging techniques to establish a final judgment. Echocardiography is the technique chosen most often since it allows for visualization of the heart in a simple, real-time and cost-effective manner. From this exam, cardiologists study the movement of the walls of the left ventricle (LV) seeking for abnormal contractions which appear immediately after the onset of ischemia. To further evaluate cardiac functionality, parameters such as blood ejection fraction are also calculated from the echocardiogram. In order to achieve these tasks, segmentation of the LV may be necessary.

Manual segmentation is a tedious and time consuming process, more so given the amount of echo exams that cardiologists perform daily due to the high prevalence of cardiac disease. In addition, due to the noise and operator-dependency inherent to ultrasound imaging, LV segmentation and MI diagnosis present high intra and inter-observer variability.  For these reasons, diagnostic approaches based on artificial intelligence are widely investigated to obtain automatic evaluation of heart functionality from echocardiography exams. These algorithms can help reduce cardiologist’s workload by assisting in the interpretation of echocardiograms in a faster, robust and accurate manner. 

The purpose of this thesis project is the development of a fully automatic artificial intelligence algorithm for the early detection of MI from 2-D echocardiography. This involves first the segmentation of the LV and then, through the assessment of certain parameters, the identification of MI. Specifically, for the first step a Deep Learning model is used (U-Net), profiting from the availability of echocardiograms with manual segmentations. For the second step, different supervised Machine Learning algorithms are tested together with Data Augmentation techniques, obtaining at the end the best results with the Random Forest model. 

Compared to current literature, this thesis’s approach identifies the specific segments of the LV that present infarction, using, for the construction of the model, existing clinical parameters. Additionally, the model demonstrates higher performance and generalizability in comparison to other papers. Finally, the generated segmentations and calculated parameters are intended to be presented to the cardiologist allowing for human verification of the produced diagnosis. The success of this algorithm encourages the possibility of future application in the clinical context; however, this first requires validation of the model with further data.

Objectives

Development of an automatic algorithm with artificial intelligence methods for the early diagnosis of myocardial infarction from echocardiography images.

Research Methodology

Pipeline consisting of 4 stages:

1. Preprocessing: preparation of echocardiography datasets
2. Segmentation: construction of deep learning network (U-Net CNN)
3. Post processing: morphological operations and feature extraction from obtained segmentation masks
4. Classification: construction and testing of different ML algorithms (KNN, SVM and RF). Classification of the patient as a whole as well as individual segments

Conclusions

High performance is obtained for the segmentation of the left ventricle through a Deep Learning network using two different datasets (high generalizability).
High performance is obtained for the identification of MI in patients in general and in individual myocardial segments. This through machine learning classifiers trained with clinical features.

Future developments

Segmentation network that uses more than one echocardiographic view to reduce the error in the following feature extraction phase.
Model validation with more data.