Detectability of Small Low-Attenuation Lesions With Deep Learning CT Image Reconstruction: A 24-Reader Phantom Study

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Iterative reconstruction (IR) techniques in CT imaging face limitations in spatial resolution, hindering radiation dose reduction.
• Deep learning image reconstruction (DLIR) offers a potential solution to overcome these contrast-dependent resolution issues.

Purpose of the Study:

• To evaluate the low-contrast detectability and radiation dose reduction capabilities of a DLIR algorithm.
• To compare DLIR performance against filtered back projection (FBP) and hybrid IR (ASiR-V) using a human reader study and observer modeling.

Main Methods:

• A dual-phantom construct with low-contrast objects was imaged at varying radiation exposures (up to 90% reduction).
• Images were reconstructed using FBP, ASiR-V, and DLIR (TrueFidelity).
• Twenty-four readers performed a two-alternative forced choice task; detectability index (d') and AUC were calculated.

Main Results:

• DLIR demonstrated noninferior low-contrast detectability compared to FBP and IR, even with up to 90% radiation reduction.
• IR was non-inferior to FBP only at 30% and 50% dose reduction.
• No significant difference in d' was found between routine dose FBP and DLIR at 70% dose reduction.

Conclusions:

• Certain DLIR algorithm settings achieve noninferior low-contrast detectability at significantly reduced radiation levels.
• DLIR shows potential for dose reduction while preserving image quality for detecting hypoattenuating lesions.
• DLIR addresses a key limitation of current IR techniques in medical imaging.