Deep-learning-based model observer for a lung nodule detection task in computed tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Task-based image quality assessment with model observers (MOs) aids CT optimization but struggles with anatomical complexity.
• Traditional MOs are limited to simplified tasks and phantoms, not complex backgrounds.
• Anatomical variability impacts human diagnostic performance in CT imaging.

Purpose of the Study:

• Develop and validate a deep-learning-based MO (DL-MO) for realistic localization tasks.
• Assess DL-MO performance in a lung nodule detection task.
• Compare DL-MO performance against radiologist readers.

Main Methods:

• A DL-MO was developed for localization tasks using projection-based lesion/noise insertion.
• DL-MO performance was compared to 4 radiologists across 12 conditions (dose, nodule size/type, reconstruction).
• DL-MO was trained on small image regions and tested with a confined search volume for efficiency.

Main Results:

• A strong correlation (Pearson's r=0.980) was found between DL-MO and human readers.
• The average performance bias between DL-MO and radiologists was minimal at 0.57%.
• The DL-MO demonstrated robust performance across various experimental conditions.

Conclusions:

• The proposed DL-MO shows potential for diagnostic image quality assessment in chest CT.
• DL-MO can accurately mimic radiologist performance in complex clinical tasks.
• This approach facilitates radiation dose and scanning protocol optimization in CT imaging.