1 Left Ventricle (LV) Segmentation using K-means and Deep Learing and volumetric calculation Azadeh Hadadi 1 , Adeyemi Abdulnafiu Olalekan 1 , Alain Lalande 2 , 1 Condorcet University Centert, Universite Bourgogne et Franche-Compte, 71200 Le Creusot, France 2 Le2I - Laboratoire de Biophysique, Facult´ e de M´ edecine Universit´ e de Bourgogne, 21079 Dijon, France The paper focuses on the automatic cardiac diagnostic challenge (ACDC) and more specifically on LV segmentation in MRI images. Two fully automatic segmentation methods, i.e., k-means and deep learning, will be presented, and discussed in detail. The implementation of each method will be described and both methods will be unified in a GUI tool for better usability. The tool is a user interface application which developed in Python and wrapped under MATLAB to give better access for further development. The objective of the tool is to segment the LV during diastole and systole (only the endocardium) in MRI images and calculate ventricle cavity volume and ejection fraction. The methods were tested on MRI images of a database made of 100 patients. The experimental evaluation demonstrates promising results and significant precision for deep learning algorithm on entire database while k-means might prone to failure if intensity changes significantly. The deep learning algorithm using TensorFlow backbone was trained several times with different training dataset and the results were compared. The findal findings shows that if the number of the images used in training increases and the diversity of the selected images covers broad specturm of the scenarios then the precsion in practice increases significantly comparing other LV segementation methods. The result shows 50% accuracy improvement when database size increases four times. Index Terms—Deep Learning, K-means, fully autmatic segmenation, Left Ventricle, MRI, cardiac diagnostic challenge. I. I NTRODUCTION Cardiovascular diseases cause 17.5 million deaths every year in all around the world [1]. Cardiac MRI image se- quences, covering of one full period of cardiac cycle or over several periods,is for evaluating cardiac function. Evaluation of cardiac function requires calculation of different cardiac pa- rameters (i.e. ejection fraction (EF), left ventricle mass (LVM), left ventricle volume, wall thickness, or wall thickening). All of these parameters can be acquired from segmented endo- cardium. Traditionally, cardiologists do manual segmentation of these contours in all dataset (i.e. all time frames in all slices) which is a very time consuming task. Therefore, the automatic segmentation of the LV is considered as a challenging task [2]. Several semi-automatic and automatic algorithms were proposed. Definitely automatic algorithm attracts more atten- tion although due to more complicated scenarios in practice achieving high precision is usually hard. Recently, repaid development in deep learning has motivated many researchers to use this approaches. Therefore this research will focus on using deep learning for the LV segmentation too. this paper will present a tool to segment the LV in diastole and systole only the endocardium using dataset aquired from https://www.creatis.insa-lyon.fr/Challenge/acdc/ and calcu- late the ventricular cavity volume in diastole and systole. for this purpose, we will set up an easy training dataset generation, training and evaluation procedure which allows not only to be used in the LV segmentation but also can be repeated on other cardiac segmentation (or other organs) and analysis using a Mask-RCCN [3]. The Mask-RCCN employs TensorFlow backbone to train a given model of conventional neural network for detection and segmentation. Different Manuscript submited May 22, 2020 for evaluation. Corresponding author: Azadeh Hadadi/Adeyemi Abdulnafiu Olalekan. training dataset and parameters lead to different procession. However, the higher precision needs more time which can go beyond few days even. In this work, we developed a tool which simplified data selection and make procedure and evaluation fully automatic with minimum user intervention. Beside, an existing successfully tested CNN architecture will be used to get training weights (gain, bias and so on) which significantly decrease to whole CNN training time. It has been practically proved that our proposed tool will help the user and researchers as well as developers to save a lot time and fairly high precision. The remaining part of the report is organized as follows: In section II, literatures of the related works will be reviewed and summarized. then, LV structure and cardiac MRI will be explained. Section III will be dedicated to selected al- gorithm and theory which briefly presents different parts of the algorithm. Section IV will detail implementation, dataset preparation, training, evaluation for both K-means and Deep learning. Results will be presented in section V and discussed. The paper will end up to conclusion and references. II. REVIEW AND RELATED WORKS A. Literature Review All algorithms presented for the LV segmentation can be divided into two categories, i.e., semi-automatic and automatic. Here in this literature only automatic segmentation will be reviewed. Hisham et al. [4] proposed a novel deep learning approach for the automated segmentation and quantification of the LV from cardiac cine MRI images using two consecutive F-CNNs, the first network for localization of ROI and the second network for precise segmentation. Dong et al. [5] uses two-networks architecture as [4] to segment the LV directly in 3D. They introduced a new network for the second stage,