Harmonizing Reporting and Identification of Lesions in Chest X-rays - Issues and their Implications for Development of an AI Tool

Abstract

This commentary addresses the critical challenges in harmonizing the reporting and identification of lesions
in chest X-rays (CXRs), particularly in the context of developing artificial intelligence (AI) tools for radiological
interpretation. Despite the significant advancements in medical imaging, the error rate in clinical practice
remains alarmingly high, with millions of misinterpretations occurring annually. AI models have shown promise in
interpreting CXRs, yet they struggle with assessing cardiomegaly, hilar abnormalities, and diaphragm positioning
in view of the subjective decision various radiologists/physicians which affects the training of the AI tool. This
paper highlights four primary issues: the complexities of accurately diagnosing cardiomegaly due to projection
variations and anatomical confounders; the challenges in recognizing hilar abnormalities due to the intricate
anatomy and variability among individuals; and the difficulties in assessing diaphragm position and shape,
which can be influenced by various physiological factors or anatomical variations. The authors advocate for the
establishment of standardized uniform objective criteria for abnormality identification in CXRs, which would
enhance the accuracy of AI models and improve clinical diagnosis. By fostering collaboration among radiologists/
physicians and AI developers, the goal is to create a uniform criteria for abnormality detection which would help
in development of more reliable diagnostic tool would minimize errors and maximize patient safety.