Scholarlink

Chest radiographs (CXR) play a vital role in the diagnosis of pulmonary disorders. However, the interpretation of CXR is generally based on stereotypical data that may lead to interobserver variability in the diagnosis. In this paper, we present the comparison of pre-trained deep learning (DL) models that are far beyond other models and were trained to automatically diagnose pulmonary disorders from CXR. Those models include a detailed comparison of ResNet50, VGG16, and Inception V3. By utilizing the 2 large datasets of radiograph images of the chest incorporating multiple symptomologies and class labels pre-clinical data of CXR images, we trained the model to utilize the functioning of the Convolution Neural Network (CNN). Utilizing the dataset, we trained the model to identify 9 different minor and major characteristics that hint at the existence of a range of pulmonary disorders such as pneumonia, tuberculosis chronic obstructive pulmonary disease (COPD), healthy states, etc. We use a multi-class classification system to sufficiently categorize CXR images and present the diagnosis with high precision. The addition of data augmentation mechanisms commonly employed to create a caption or notice significant areas in CNN to the proposed model assists in focusing extra interest on abnormal and crucial areas within CXR images for better interpretation. Extensive experiments were undertaken to demonstrate that the model is effective and superior to all modern rivals concerning performance. Additionally, comprehensive assessment and validation were carried out to conclude the solidity and essential applicability qualities in various patient-only datasets and image acquisition environments. Our experiments show the best results in ResNet50 with an accuracy of 97.36%. To summarize, as a contribution rather than a novelty in DL, we showed that pre-trained models can offer a great opportunity to increase the amount of feasibility in pulmonary disorder diagnoses. Moreover, these models can be used in automated systems to detect pulmonary disorders effectively.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Deep Learning Models for Chest Radiograph Classification: A Multi-model Evaluation with Image Enhancement Techniques to Diagnose Pulmonary Disorders

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Iftekhar Alam Sarker,
  • Ab. Rahim Ahmed Sowrov,
  • Sadia Islam Prova,
  • Md. Adnan Morshed,
  • Ahmed Wasif Reza

摘要

Chest radiographs (CXR) play a vital role in the diagnosis of pulmonary disorders. However, the interpretation of CXR is generally based on stereotypical data that may lead to interobserver variability in the diagnosis. In this paper, we present the comparison of pre-trained deep learning (DL) models that are far beyond other models and were trained to automatically diagnose pulmonary disorders from CXR. Those models include a detailed comparison of ResNet50, VGG16, and Inception V3. By utilizing the 2 large datasets of radiograph images of the chest incorporating multiple symptomologies and class labels pre-clinical data of CXR images, we trained the model to utilize the functioning of the Convolution Neural Network (CNN). Utilizing the dataset, we trained the model to identify 9 different minor and major characteristics that hint at the existence of a range of pulmonary disorders such as pneumonia, tuberculosis chronic obstructive pulmonary disease (COPD), healthy states, etc. We use a multi-class classification system to sufficiently categorize CXR images and present the diagnosis with high precision. The addition of data augmentation mechanisms commonly employed to create a caption or notice significant areas in CNN to the proposed model assists in focusing extra interest on abnormal and crucial areas within CXR images for better interpretation. Extensive experiments were undertaken to demonstrate that the model is effective and superior to all modern rivals concerning performance. Additionally, comprehensive assessment and validation were carried out to conclude the solidity and essential applicability qualities in various patient-only datasets and image acquisition environments. Our experiments show the best results in ResNet50 with an accuracy of 97.36%. To summarize, as a contribution rather than a novelty in DL, we showed that pre-trained models can offer a great opportunity to increase the amount of feasibility in pulmonary disorder diagnoses. Moreover, these models can be used in automated systems to detect pulmonary disorders effectively.