MRI Cardiac Images Segmentation and Anomaly Detection Using U-Net Convolutional Neural Networks

Abstract

Healthcare industry is increasingly adopting artificial intelligence in analyzing laboratory and radiology outputs to provide optimal treatments for patients. Medical imaging has been playing an important role in understanding several underlying conditions of patients. With rise in incidents of cardiac diseases world-wide, usage of computer vision and deep learning methods are proving to be very useful in detecting anomalies that are conventionally done using human perception. This paper aims at establishing efficacy of using convolutional neural network in detecting cardiac anomaly. The hypothesis is substantiated through the process of predicting systole and diastole volumes from MRI scan images of left ventricle and calculation of ejection fraction which is a vital parameter in assessing cardiac dysfunction. In this research project, representative dataset and normalization techniques were used to arrive at the results published. With the progression in medical imaging techniques combined with training the model with high volume of stratified data can result in very high accuracy of outcome. The model was successfully able to predict the diastole and systole volumes of any given image based on image segmentation, pixel values, slice locations, and other DICOM metadata. The cardiac anomaly was determined using calculated ejection fraction from the predicted systolic and diastolic volumes.