Classification of Left Ventricle and Non- Left Ventricle Segment for Cardiac Assessment Using Deep Convolutional Neural Network Article Swipe
YOU?
·
· 2022
· Open Access
·
· DOI: https://doi.org/10.24191/jeesr.v21i1.005
· OA: W4308262305
In large-scale medical imaging, selecting the best image to extract relevant imaging biomarkers for image assessment is crucial.Segmentation of the left ventricle (LV) and myocardium are performed in computer-aided analysis usually at short-axis slices of cardiac magnetic resonance (MR) image to quantify cardiovascular disease assessment, such as myocardial scarring, LV ejection fraction and LV mass.The need to correctly identify a short-axis slice range for efficient quantification is preferred for automatic classification of the slice range of interest.The goal of this research is to establish an image processing method for the segmentation of Left ventricle scar from late gadolinium-enhanced (LGE) MR images.In order to achieve the main purpose, the work is divided into two parts, the first is to identify the cardiac Left ventricle segment (LVS) in the stack of short-axis LGE MR images and the second part; detecting the scar in between the LV myocardium area.This paper will present the outcomes of the first part by utilizing a deep convolutional neural network (DCNN) to construct an automatic system for classifying LVS and Non-Left Ventricle Segment (Non-LVS) in MR images.The same image dataset will be used for a comparative analysis with six DCNN models designed from scratch and three famed pre-trained networks, Alexnet, GoogleNet and SqueezeNet.Each model is trained up to 35 epochs using the Cardiac Atlas dataset and cross-validation method.The outcome from this work demonstrated that the DCNN3Y performs well over small training data with an average accuracy of 94.49%.Whereas SqueezeNet outperformed the three pre-trained networks with an average accuracy of 96.96%.It has also been discovered that increasing the number of filters and their subsequent configuration slightly influences the network's performance.This produces very promising results showing that it is ready to be used in the second part of this research.The outcome of this research can compensate for the deficiencies of manual detection in the original image detection system, increase detection efficiency, reduce detection misjudgments, and advance the development of automated and intelligent detection in the medical field.
