Correlation of Automated in Vivo Image Quality With Radiologist's Performance in Abdomen Computed Tomography Across...

Area of Science:

• Radiology
• Medical Imaging
• Computational Imaging

Background:

• Radiologists' performance is a key indicator of diagnostic image quality.
• Evaluating image quality requires methods that correlate with clinical performance.
• Current metrics may not adequately reflect performance with advanced reconstruction algorithms.

Purpose of the Study:

• To assess the correlation between an in vivo-characterized detectability index (d') for liver lesions in CT and radiologists' performance.
• To compare this correlation across two different CT reconstruction algorithms: Filtered Back Projection (FBP) and deep learning (DL).
• To evaluate a new formalism (adj) for image quality measurement that accounts for DL reconstruction non-linearity.

Main Methods:

• Fifty-one contrast-enhanced abdominal CT studies for colorectal liver metastases were analyzed.
• Images were reconstructed using FBP and DL algorithms.
• Expert radiologists performed lesion detection and malignancy likelihood assessments.
• Task-based performance was evaluated using conventional d' and the new adjusted d' (adj) formalism.

Main Results:

• Conventional d' correlated well with radiologist performance for FBP but not DL images.
• The new adj formalism showed consistent reflection of performance across both FBP and DL reconstructions.
• For small lesions (<=6 mm), adj differences were smaller (-9%) compared to conventional d' (34%) when comparing DL to FBP.
• For medium lesions (6-10 mm), adj differences were -13% vs. 29% for conventional d'.

Conclusions:

• The new adj formalism robustly reflects CT system clinical performance regardless of the reconstruction algorithm used.
• This methodology provides a more accurate assessment of real-world CT system performance.
• The adj metric is a valuable tool for evaluating advanced CT imaging techniques.