Low-contrast lesion detection in neck CT: a multireader study comparing deep learning, iterative, and filtered back...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Computed Tomography (CT) reconstruction algorithms enhance image quality.
• Deep Learning Reconstruction (DLR) shows promise for improving CT image quality.
• Comparison of DLR, Iterative Reconstruction (IR), and Filtered Back Projection (FBP) is crucial for neck CT lesion detection.

Purpose of the Study:

• To compare the efficacy of DLR, IR, and FBP in detecting and localizing lesions in neck CT.
• To evaluate the impact of varying radiation doses on lesion detection across different reconstruction algorithms.
• To determine if DLR can maintain diagnostic performance at reduced radiation doses.

Main Methods:

• Nine neck phantoms with simulated lesions were scanned at six different radiation doses (0.5–5.2 mGy) using a 320-slice CT scanner.
• Images were reconstructed using FBP, IR, and DLR algorithms.
• Thirteen readers assessed lesion detectability and localization using Receiver Operating Characteristic (ROC) and Localization ROC (LROC) analyses.

Main Results:

• DLR demonstrated superior lesion detection (ROC AUC 0.724) compared to IR (0.696) and FBP (0.671).
• DLR also showed improved lesion localization (LROC AUC 0.407) over IR (0.338) and FBP (0.313).
• Low-dose CT (0.5 mGy) with DLR or IR maintained lesion detection, unlike FBP where detection was compromised.

Conclusions:

• Deep learning reconstruction significantly enhances lesion detection and localization in neck CT imaging.
• DLR offers improved diagnostic performance compared to IR and FBP.
• DLR facilitates effective radiation dose reduction while preserving lesion detectability.