Robustness of Deep Learning Algorithm to Varying Imaging Conditions in Detecting Low Contrast Objects In Computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computer-Aided Diagnosis

Background:

• Radiologist performance in detecting low-contrast objects in CT images can vary.
• Evaluating new algorithms requires comparison against human expert performance.

Purpose of the Study:

• To compare the diagnostic performance of a deep learning algorithm (DLA) against radiologists for detecting low-contrast objects in CT phantom images.
• To assess the robustness of the DLA under various imaging conditions.

Main Methods:

• A deep learning algorithm (DLA) was trained on 10,000 synthetic CT phantom images with varying object sizes and contrast differences.
• Twelve radiologists evaluated 640 real CT phantom images (Catphan®) using a five-point scale.
• Performance was quantified using area under the receiver operating characteristics curve (AUC) and compared between the DLA and radiologists.

Main Results:

• The DLA achieved a consistently higher AUC than radiologists across all tested imaging conditions (p < 0.0001).
• The DLA exhibited significantly less performance degradation under varying imaging conditions compared to radiologists.
• Post-hoc analysis using bootstrapping confirmed the DLA's superior and more robust performance.

Conclusions:

• Deep learning algorithms can outperform radiologists in detecting low-contrast objects in CT imaging.
• DLAs offer a more robust and consistent diagnostic performance, particularly under challenging or variable imaging conditions.
• This suggests a significant potential for DLAs to augment or improve radiological detection tasks.