The Role of Artificial Intelligence in Optimizing X-Ray Image Quality and Reducing Radiation Dose: A Retrospective Comparative Study

Abstract

Background: X-rays (plain film / conventional radiography) remain the most used diagnostic imaging technique worldwide and radiation dose optimization and image quality remain as basic principles of radioprotection. Imaging optimization using deep-learning image processing to achieve comparable image quality at reduced dose has been suggested. Purpose: This study was conducted to evaluate the difference in radiation dose, the objective quality of the image, subjective image quality, and diagnostic performance between a conventional-dose radiography protocol and an AI-optimized reduced dose protocol.
Materials and Methods: In this retrospective, comparative study, 300 X-ray examinations (chest, musculoskeletal and abdominal) were included, three each were randomly assigned to a control group (standard dose, n = 150) and experimental group (optimized low-dose; n = 150). Dose measures (kV, mAs, DAP, ESD, ED), objective measurements (SNR, CNR, noise, edge sharpness index) and 5-point Likert scores of two radiologists were compared. Subgroup and ROC analyses were performed.
Results: Baseline characteristics were similar (all p > 0.50). Effective dose was reduced 54.9% with the AI-optimized protocol (0.082 ± 0.015 vs 0.037 ± 0.008mSv, p < 0.001) and dose-area product by 53.9% and signal-to-noise by 25.4%, contrast-to-noise by 34.6%, and noise by 35.7% (all p < 0.001). Radiologist ratings improved with AI for all criteria (p < 0.001) with inter-observer reliability 0.78-0.89. No differences in diagnostic acceptability (99.3% vs 96.7%, p = 0.121) or area under the curve (0.982 vs 0.964, p = 0.118) were observed. Effects did not vary between chest and musculoskeletal scans (p-interaction > 0.05).
Conclusion: In this retrospective cohort, AI-assisted optimization resulted in an approximately 55% decrease in radiation dose, while maintaining or even improving image quality, with no objectively proven loss in diagnostic performance. Multicenter prospective validation is required before general implementation.