Enhanced Deep Learning Model for Pneumonia Segmentation in Chest X-Ray Images

Pneumonia from images of chest X-rays requires accurate segmentation for effective diagnosis and treatment. This study proposes an enhanced deep learning model, EASUNet, designed to integrate the EfficientNet-B4 backbone, Atrous Spatial Pyramid Pooling (ASPP) and a U-Net decoder with attention mechanisms to refine segmentation accuracy. The methodology preprocesses ChestX-ray14 data, uses the EfficientNet-B4 backbone for capturing hierarchical features, ASPP module to focus on different scales and contexts in the X-ray images and the U-Net decoder, augmented with attention mechanisms to prioritize relevant features and suppress irrelevant areas improving the model's focus on pneumonia-affected regions. Comparative results show that EASU-Net outperforms traditional models like U-Net, Attention U-Net and DeepLabV3 by achieving a Dice Coefficient of 0.93, accuracy of 95%, and Intersection over Union (IoU) of 0.88. The results suggest that EASU-Net is a reliable segmentation tool, providing medical professionals with improved precision for diagnosing pneumonia from chest X-rays.