Improved Segmentation of Cardiac MRI Using Efficient Pre-Processing Techniques

Segmentation Techniques

LV Segmentation is a tedious and hard task and numerous techniques have been developed for the same. Vincent et al. (1991) suggested a watershed algorithm which combines region merging and thresholding. The image gradient map is found and threshold is set on the image gradient’s magnitude. Some methods have also been suggested which directly estimate the volume of the right and left chamber of the heart, without performing the process of segmentation (Ashfin et al. 2014; Wang et al. 2014; Zhen et al. 2015). Single atlas based as well as multi atlas-based segmentation approaches have been applied on cardiac MR images (Heckemann et al. 2006; Artaechevarria et al. 2009; Sabuncu et al. 2010; Warfield et al. 2014; Asman & Landman 2011).

Lee et al. (2009) proposes LV segmentation method by using Iterative Thresholding method and deep convolutional encoder-decoder model. Tran (2017) makes use of a fully convolutional neural network for segmentation. Medical images have also been segmented by using the U-Net architecture (Ronneberger et al. 2015). The accuracy of segmentation has been improved by removing the uncertainty of deep neural network (Norouzi, 2019). LV segmentation is used to analyse the blood flow in the heart and the segmentation is improved by making use of intramodality image registration (Gupta et al. 2018). Luo et al. (2018) uses hierarchical extreme learning machine model for performing segmentation of the LV. ZhenZhou (2016, 2017) has suggested remarkable work in the field of LV segmentation.

State-of-the-art segmentation techniques include the use of fully convolution network which perform semantic segmentation (Long et al. 2015) and various modifications have also been performed to it (Garcia et al. 2017). Various researchers have used machine learning algorithms in combination with deformable models (Ngo et al. 2013). Dynamic programming approach has also been used for fast segmentation of cardiac MRI (Santiago et al. 2017). Chen et al. (2020) presents a review of cardiac image segmentation methods which make use of deep learning. Deep learning has its own challenges as well. Mahony et al.(2019) has provided the limitations that are faced by deep learning methods in comparison to the traditional computer vision techniques. Deep learning methods require high computational cost and a strong graphics processing unit for training the model. In this paper A novel approach has been proposed in this paper which increases the segmentation accuracy without making use of deep learning methods.