Added Value of Deep Learning-based Detection System for Multiple Major Findings on Chest Radiographs: A Randomized...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Previous computer-aided detection studies on chest radiographs may have biased results due to sequential reading designs.
• Observer performance in detecting major radiographic abnormalities requires objective evaluation.

Purpose of the Study:

• To compare observer performance in detecting and localizing major abnormal findings on chest radiographs with and without deep learning-based detection (DLD) system assistance.
• To utilize a randomized crossover design to minimize bias in performance assessment.

Main Methods:

• Retrospectively collected chest radiographs (normal and abnormal) from 2016-2017 were used.
• Six observers interpreted radiographs with and without DLD system assistance in a randomized crossover design.
• Performance metrics included JAFROC FOM, AUC, sensitivity, specificity, and reading times.

Main Results:

• DLD system use significantly improved JAFROC FOM (0.90 to 0.95) and AUC (0.93 to 0.98).
• Per-lesion sensitivity increased from 83% to 89.1%, per-image sensitivity from 80% to 89%, and specificity from 89.3% to 96.6%.
• Reading time was reduced from 10-65 seconds to 6-27 seconds with DLD assistance.

Conclusions:

• Deep learning-based detection systems enhance observer performance in detecting and localizing major abnormal findings on chest radiographs.
• The DLD system demonstrated superior performance compared to pooled observers.
• AI assistance leads to improved diagnostic accuracy and efficiency in chest radiograph interpretation.