Khokan Md Ibrahim Patwary, Tonni Tasnim Jahan, Rony Md Awlad Hossen, Fatema Kaniz, Hasan Md Zahid
Health Informatics Research Lab, Department of Computer Science and Engineering, Daffodil International University, Dhaka, 1216, Bangladesh.
Comput Biol Med. 2025 Sep;195:110588. doi: 10.1016/j.compbiomed.2025.110588. Epub 2025 Jun 25.
Respiratory disorders cause approximately 4 million deaths annually worldwide, making them the third leading cause of mortality. Early detection is critical to improving survival rates and recovery outcomes. However, chest X-rays require expertise, and computational intelligence provides valuable support to improve diagnostic accuracy and support medical professionals in decision-making. This study presents an automated system to classify respiratory diseases using three diverse datasets comprising 18,000 chest X-ray images and masks, categorized into six classes. Image preprocessing techniques, such as resizing for input standardization and CLAHE for contrast enhancement, were applied to ensure uniformity and improve the visual quality of the images. Albumentations-based augmentation methods addressed class imbalances, while bitwise segmentation focused on extracting the region of interest (ROI). Furthermore, clinically handcrafted feature extraction enabled the accurate identification of 20 critical clinical features essential for disease classification. The K-nearest neighbors (KNN) graph construction technique was utilized to transform tabular data into graph structures for effective node classification. We employed feature analysis to identify critical attributes that contribute to class predictions within the graph structure. Additionally, the GNNExplainer was utilized to validate these findings by highlighting significant nodes, edges, and features that influence the model's decision-making process. The proposed model, Chest X-ray Graph Neural Network (CHXGNN), a robust Graph Neural Network (GNN) architecture, incorporates advanced layers, batch normalization, dropout regularization, and optimization strategies. Extensive testing and ablation studies demonstrated the model's exceptional performance, achieving an accuracy of 99.56 %. Our CHXGNN model shows significant potential in detecting and classifying respiratory diseases, promising to enhance diagnostic efficiency and improve patient outcomes in respiratory healthcare.
呼吸系统疾病每年在全球导致约400万人死亡，使其成为第三大死亡原因。早期检测对于提高生存率和康复效果至关重要。然而，胸部X光检查需要专业知识，而计算智能为提高诊断准确性和支持医疗专业人员进行决策提供了宝贵的支持。本研究提出了一个自动化系统，用于使用三个包含18000张胸部X光图像和掩码的不同数据集对呼吸系统疾病进行分类，这些数据集分为六个类别。应用了图像预处理技术，如调整大小以进行输入标准化和使用对比度受限自适应直方图均衡化（CLAHE）进行对比度增强，以确保图像的一致性并提高视觉质量。基于数据增强的方法解决了类别不平衡问题，而逐位分割则专注于提取感兴趣区域（ROI）。此外，临床手工特征提取能够准确识别疾病分类所需的20个关键临床特征。利用K近邻（KNN）图构建技术将表格数据转换为图结构，以进行有效的节点分类。我们采用特征分析来识别在图结构中有助于类别预测的关键属性。此外，利用GNNExplainer通过突出影响模型决策过程的重要节点、边和特征来验证这些发现。所提出的模型，即胸部X光图神经网络（CHXGNN），是一种强大的图神经网络（GNN）架构，它结合了先进的层、批量归一化、随机失活正则化和优化策略。广泛的测试和消融研究证明了该模型的卓越性能，准确率达到了99.56%。我们的CHXGNN模型在检测和分类呼吸系统疾病方面显示出巨大潜力，有望提高呼吸医疗保健中的诊断效率并改善患者预后。
