Identity-Mapping ResFormer: A Computer-Aided Diagnosis Model for Pneumonia X-Ray Images

作     者：Zhou, Tao Peng, Caiyue Guo, Yujie Wang, Hongxia Niu, Yuxia Lu, Huiling 

作者机构：North Minzu Univ Sch Comp Sci & Engn Yinchuan 750021 Peoples R China North Minzu Univ Key Lab Image & Graph Intelligent Proc State Ethn Yinchuan 750021 Peoples R China Ningxia Med Univ Sch Med Informat & Engn Yinchuan 750004 Peoples R China 

出 版 物：《IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT》 (IEEE Trans. Instrum. Meas.)

年 卷 期：2025年第74卷

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0804[工学-仪器科学与技术] 

基　　金：National Natural Science Foundation of China Ningxia Natural Science Foundation Project [2023AAC03293] 

主　　题：Feature extraction Convolution Lungs Pneumonia Transformers X-ray imaging Computational modeling Shape Lesions Residual neural networks Cross-scale fusion identity-mapping pneumonia computer-aided diagnosis residual neural network (ResNet) transformer 

摘      要：Pneumonia is a serious threat to human health and causes great harm to the human respiratory system. Pneumonia recognition based on lung X-ray images is an effective auxiliary diagnosis method. In the lung X-ray images, the lesion areas are complex and diverse, and the boundary is not clear. There is a gradient inconsistency problem in the feature interaction process of deep learning. They prevent the model from focusing well on diseased areas in the image. To solve these problems, this article proposes a computer-aided diagnosis model for pneumonia X-ray images-identity-mapping ResFormer. The main innovations are as follows: First, the multiconvolution cascade residual module (MCCRM) is designed to extract local image features of different sizes and shapes. The MCCRM is a parallel-cascade structure, which enhances the feature extraction capability of the backbone network by nesting four convolution operations. Second, the enhanced multipatch transformer (EMPT) is designed in the auxiliary network to extract the multiperceptive field s global attention features. It enables the network to focus on prominent area features of the image. Third, the identity-mapping transformer module (IMTM) is designed to solve the gradient inconsistency problem in different stage features. Transformer operations are used in this module to fuse gradient features in different stages. Finally, the model is validated on a lung X-ray dataset. The accuracy, F1 , recall, precision, and specificity are 97.6679%, 95.3378%, 95.3368%, 95.4602%, and 98.4452%, respectively. Identity-mapping ResFormer can assist doctors to make efficient and accurate pneumonia diagnoses.