The Effect of Simulated Dose Reduction on the Performance of Artificial Intelligence in Chest Radiography

Area of Science:

• Medical imaging
• Artificial intelligence in radiology
• Radiography

Background:

• Chest imaging is vital for patient screening and monitoring.
• Artificial intelligence (AI) models are increasingly used to support diagnostic tasks in chest radiography.
• The impact of reduced radiation dose and image quality on AI performance remains largely unexplored.

Purpose of the Study:

• To simulate low-dose chest radiography images.
• To evaluate the effect of simulated dose reduction on the performance of Convolutional Neural Networks (CNNs).
• To compare the impact of dose reduction versus patient positioning on AI model predictions.

Main Methods:

• Trained 14 CNN models using seven pathology labels from the MIMIC dataset at two spatial resolutions.
• Tested models on original, 50%, and 75% low-dose simulated images.
• Compared Area Under the Receiver Operator Characteristic (AUROC) curves using DeLong testing, analyzing results by gender and patient position.

Main Results:

• Average absolute change in AUROC was <0.005 for simulated dose reductions, with a maximum change of 0.014.
• Significant differences in AUROC were observed in 6 out of 28 test sets.
• Patient positioning demonstrated a more substantial effect on CNN performance than simulated dose reduction.

Conclusions:

• Simulated dose reduction in chest radiography has a minimal clinical impact on CNN performance.
• While statistically significant differences exist in some cases, the overall effect of dose reduction is minor.
• Patient positioning is a more critical factor influencing AI performance in chest radiography than radiation dose reduction.