Date of Publication :30th November 2023

Abstract: Over the recent years, chest diseases have become a severe problem that has a significant impact on people’s lives all around the world. If these diseases are not detected in time, they can be fatal and cause death. Chest radiography (CXR) is a cost-effective method of identifying and diagnosing diseases. However, identifying chest anomalies from CXR images is time-consuming and needs professional radiologists. Automatic biomedical image segmentation helps speed up disease detection and diagnosis. So, a fully automated system is necessary, and there is a lot of work regarding this. Rapid advancements in deep learning have produced groundbreaking outcomes in this area. Modern networks like U-Net and SegNet, however, frequently perform poorly in difficult domains. This paper proposes a fully automated system for chest disease detection along with a segmentation module. We have combined the classification and segmentation tasks into a single model and have developed a novel architecture that is based on Nested UNet architecture. Besides, this work also quantifies the infection rate of a chest and localizes the infected region of the chest due to COVID-19. The proposed approach uses a transfer learning enriched pre-trained encoder to learn enough from the limited amount of data. It has modified skip connections to reduce semantic gaps between the corresponding levels of the encoder-decoder layer. The proposed architecture has comparatively fewer parameters than most of the recent works. As a result, it has a lightweight design and is less likely to overfit. With an iterative procedure, the created network improved lung area segmentation performance, achieving an Intersection over Union (IoU) of 93.59% and a Dice Similarity Coefficient (DSC) of 97.61%. Additionally, with 97.67% Intersection over Union (IoU) and 87.61% DSC, COVID-19 infections of varied shapes were perfectly localized. Finally, the proposed model has obtained an excellent chest disease detection performance with an accuracy of 92.86% which is satisfying. The remainder of the study outlines a novel method for automating the detection and segmentation of chest diseases from CXR pictures using deep learning, transfer learning, and a special Nested U-Net architecture.

Reference :