Artificial Intelligence in Chest Radiography Comparative Study with Hospital Radiologists’ Reports in Dhaka, Bangladesh

Abstract

Artificial intelligence (AI) has emerged as a promising tool in radiology, particularly chest radiography, where timely and accurate diagnosis is critical for patient care. This study aimed to compare the diagnostic performance of a deep learning model, DenseNet-121, with radiologist reports of chest X-ray (CXR) images in a hospital setting in Dhaka, Bangladesh. A total of 50 posteroanterior (P/A) chest X-rays were collected from a single hospital and analyzed independently using the DenseNet-121 algorithm and radiologists. The AI system reported 42 images as normal and eight as abnormal, identifying findings such as pneumonitis, pleural effusion, cardiomegaly, lung opacities, nodules, and tracheal shift. In contrast, radiologists classified 37 images as normal and 13 as abnormal using standard reporting formats that assessed lung fields, heart, basal angles, diaphragm, bony thorax, and tracheal position. The comparative analysis revealed areas of concordance, particularly in normal findings, where both the AI and radiologists demonstrated similar interpretations. However, discrepancies were noted in the abnormal cases. While AI occasionally fails to localize specific pathologies, radiologists also miss abnormalities in a few instances, possibly due to workload or reporting variations. These differences underscore the complementary role of AI, which, despite requiring further training and refinement, can provide significant support for diagnostic accuracy and efficiency. Overall, this study highlights the potential of DenseNet-121 as an assistive diagnostic tool for chest radiography in resource-constrained health care settings. By enhancing radiologists’ confidence and reducing diagnostic oversight, AI can contribute to improved clinical decision making and patient outcomes in Bangladesh and beyond.

Ähnliche Arbeiten