AI-Based Radiological Diagnosis in Pulmonary Disorders

Abstract

Pulmonary disorders -- pneumonia, tuberculosis, lung cancer, chronic obstructive pulmonary disease, and interstitial lung disease -- collectively represent one of the leading causes of death globally, and chest radiography remains the first-line imaging modality for their detection. Yet radiological interpretation is surprisingly inconsistent: inter-reader agreement among radiologists for key findings on chest X-rays typically hovers around kappa 0.40-0.60, and missed diagnoses on chest radiographs are among the most common sources of diagnostic error in medicine. Artificial intelligence, particularly deep convolutional neural networks, has shown considerable promise in automating and improving radiological diagnosis, but most published studies evaluate single models on single diseases using single-institution datasets. We present the Pulmonary AI Diagnostic Framework (PAIDF), comparing five AI architectures -- standard CNNs (ResNet-50), attention-augmented CNNs (SE-ResNet), vision transformers (ViT-B/16), hybrid CNN-transformers (CoAtNet), and multi-task ensemble models -- across four diagnostic tasks: pneumonia detection, tuberculosis screening, lung nodule classification, and multi-label pulmonary disease classification. We used 148,320 chest radiographs from three public datasets (CheXpert, MIMIC-CXR, and the Shenzhen TB set) plus 12,640 radiographs from two European clinical sites (Zurich and Tallinn). Our Radiological AI Performance Score (RAPS) integrates diagnostic sensitivity, specificity, calibration quality, robustness to distribution shift, and inference speed. The multi-task ensemble achieved the highest RAPS (0.934) through complementary feature extraction across architectures, while the vision transformer achieved the highest single-model sensitivity for lung nodule classification (0.952). These findings demonstrate that ensemble approaches combining CNNs and transformers offer the most reliable path toward clinical deployment of AI-based pulmonary diagnosis