Automatic Left Ventricle Segmentation from Short-Axis MRI Images Using U-Net with Study of the Papillary Muscles’ Removal Effect

Purpose In clinical routines, the evaluation of cardiac wall motion is based on manual segmentation of ventricular contours. This task is time-consuming and leads to inter-observer variability. In this context, the aim of this paper is to propose a fully automatic method based on U-net architecture for left ventricle (LV) segmentation while studying the impact of papillary muscles presence and elimination on the segmentation accuracy. Methods In this work, we developed and evaluated an automatic approach based on U-Net architecture for LV segmentation. We started with a preprocessing pipeline which consists in cropping original images using convolutional neural network (CNN) and eliminating pillars using morphological operators. Regarding segmentation, our neural network was trained and validated using ACDC dataset composed of 150 patients. The performance of the proposed method was evaluated on an internal database composed of 100 patients (more than 2500 frames) using technical metrics including Hausdorff distance (HD), Jaccard coefficient (IoU), and Dice Similarity Coefficient (DSC). Results A comparative study demonstrated that the proposed architecture outperformed the original U-Net. Quantitative analysis of the obtained results confirmed the strength of our method that reveals the superlative segmentation performance as evaluated using the following indices including HD = 6.541 ± 1.6 mm, IoU = 94.85 ± 2%, and DSC = 93.27 ± 5% with p value