Image segmentation using improved U-Net model and convolutional block attention module based on cardiac magnetic resonance imaging

Abstract

Automated segmentation methods for cardiac magnetic resonance imaging (MRI) offer valuable assistance in evaluating cardiac function for clinical diagnosis. Nevertheless, prevailing techniques encounter challenges in dealing with characteristics like indistinct image boundaries and uneven resolution in cardiac MRI scans. Consequently, these methods often encounter problems related to uncertainty within the same class of structures and ambiguity when distinguishing between different classes. In our paper, enhancements are made to the U-Net model to address these challenges. Initially, an enhanced residual block is incorporated into the U-Net architecture to increase network depth and capture richer feature information. Subsequently, by integrating the Convolutional Block Attention Module (CBAM) mechanism, the network focuses more intensely on specific feature layers and spatial regions. This leads to the suppression of non-target region features, consequently enhancing segmentation accuracy. This study examined an enhanced model's performance using a collection of cardiac MRI data from the Straits Affiliated Hospital of Huaqiao University. The dataset encompassed 6680 images for training and 2225 images for testing. Model evaluation was conducted using the Dice coefficient and accuracy metrics, yielding values of 0.9292 and 0.8911, respectively. The outcomes indicate that the enhanced model effectively enhances the precision and accuracy of cardiac magnetic resonance imaging segmentation. The method put forth in this study brings about a substantial enhancement in segmentation accuracy, resulting in a segmentation outcome that aligns closely with the reference ground truth labels. In comparison to alternative algorithms, this approach demonstrates elevated accuracy across distinct regions, thereby yielding segmentation results of heightened precision.

Topics & Concepts