[New Method of Paired Comparison for Improved Observer Shortage Using Deep Learning Models].

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Assessing medical image quality is crucial for accurate diagnosis.
• Human observers perform paired comparisons to evaluate image quality, which can be time-consuming.
• Deep learning (DL) offers a potential automated solution for image quality assessment.

Purpose of the Study:

• To validate the use of a deep learning observer as a substitute for human observers in paired comparisons of computed tomography (CT) images.
• To assess the performance of DL models in evaluating image quality under varying imaging conditions.

Main Methods:

• Computed tomography phantom images were acquired under six different imaging conditions with varying tube currents (20-200 mA).
• Fourteen experienced radiologic technologists performed pairwise comparisons using Ura's method.
• Deep learning models (VGG16 and VGG19) were trained and evaluated for accuracy, recall, precision, specificity, and F1-score.

Main Results:

• The deep learning model achieved an average accuracy of 82%.
• The DL model's average degree of preference showed a small average difference of 0.05 compared to the human observer standard.
• Significant differences were detected by the DL model for image pairs with tube currents of 160 mA vs. 120 mA and 200 mA vs. 160 mA.

Conclusions:

• Deep learning models show potential as observers in paired comparisons for evaluating CT image quality, particularly with limited phantoms and noise evaluations.
• AI can assist in image quality assessment, potentially improving efficiency in radiology workflows.