CNN as model observer in a liver lesion detection task for x-ray computed tomography: A phantom study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Evaluating human observer performance in medical imaging tasks is crucial for diagnostic accuracy.
• Developing computational models that mimic human performance can aid in optimizing imaging protocols and training.
• Computed tomography (CT) is a key modality for detecting liver lesions, necessitating accurate performance assessment.

Purpose of the Study:

• To evaluate anthropomorphic model observers trained with neural networks for predicting human observer performance in liver lesion detection.
• To compare the predictive accuracy of different model observers, including convolutional neural networks (CNNs), against human readers.
• To assess the impact of training strategies on model observer performance.

Main Methods:

• Simulated liver lesions using contrast-enhanced phantoms were scanned using CT.
• Image data and human reader confidence ratings were used to train anthropomorphic model observers (softmax regression and CNNs).
• Model observers were evaluated using receiver operating characteristic (ROC) analysis and compared to human performance and a channelized Hotelling observer (CHOi).

Main Results:

• High correlations (r > 0.9) were observed between model observers and human performance across lesion sizes and dose levels.
• Convolutional neural network model observers (CNN-MO) demonstrated superior accuracy, with mean absolute percentage differences (MAPD) as low as 1.2% when trained with strategy A.
• Both training strategies (A: separate models per size; B: one model for all sizes) showed strong correlations, with CNN-MO performing consistently well.

Conclusions:

• Convolutional neural network model observers can accurately predict human performance in liver lesion detection tasks.
• These AI-driven models are reliable across all evaluated lesion sizes and radiation dose levels.
• CNN-based anthropomorphic observers offer a promising tool for assessing and potentially improving diagnostic performance in radiology.