A deep learning- and partial least square regression-based model observer for a low-contrast lesion detection task in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Assessing computed tomography (CT) image quality is crucial for accurate diagnosis.
• Current methods often rely on subjective human observer (HO) performance.
• Developing objective and reliable image quality assessment tools is essential.

Purpose of the Study:

• To develop a novel deep learning-based model observer (DL-MO) framework.
• To validate the DL-MO's performance in a low-contrast lesion detection task using CT images.
• To assess CT image quality directly from patient images in clinically relevant tasks.

Main Methods:

• A DL-MO was created using transfer learning with a pretrained deep convolutional neural network (CNN) and partial least square regression discriminant analysis (PLS-DA).
• The DL-MO incorporated an internal noise component to simulate human observer variability.
• Performance was evaluated against 4 medical physicists performing a two-alternative forced choice (2AFC) detection task on CT images with varying lesion characteristics and radiation doses.

Main Results:

• A statistically significant positive correlation (Pearson's r = 0.986) was found between DL-MO and HO performance.
• Bland-Altman analysis showed no significant differences between DL-MO and HO.
• The DL-MO accurately predicted human performance in a challenging low-contrast lesion detection task.

Conclusions:

• The developed DL-MO demonstrates high correlation with human observers for CT image quality assessment.
• This DL-MO framework shows significant potential for objective and clinically relevant image quality evaluation.
• The study validates the use of DL-MO for assessing image quality directly from patient CT data.