A convolutional neural network-based anthropomorphic model observer for signal-known-statistically and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Observer Models

Background:

• Accurate prediction of human observer performance is crucial for optimizing medical imaging systems.
• Conventional anthropomorphic model observers have limitations in predicting human performance, especially with complex image backgrounds.

Purpose of the Study:

• To implement a convolutional neural network (CNN)-based anthropomorphic model observer for signal-known-statistically (SKS) and background-known-statistically (BKS) detection tasks.
• To compare the performance of the CNN model observer against conventional model observers in predicting human performance on simulated cone beam computed tomography (CBCT) images.

Main Methods:

• Simulated CBCT images with varied signals and breast anatomical backgrounds were used for SKS/BKS detection tasks.
• Conventional model observers (non-prewhitening with eye-filter, Difference-of-Gaussian CHO, Gabor CHO) and a CNN-based model observer were implemented.
• A novel data labeling strategy was developed for CNN training, reflecting human decision-making inefficiencies.

Main Results:

• A three-layer CNN, trained with the proposed labeling strategy, demonstrated superior prediction of human observer performance compared to conventional model observers.
• The CNN model observer showed better prediction accuracy across different background noise structures in CBCT images.
• The trained CNN exhibited good correlation with human performance even when training and testing images had differing noise structures.

Conclusions:

• CNN-based model observers offer a more accurate method for predicting human performance in medical image analysis tasks.
• The proposed data labeling strategy enhances the effectiveness of CNNs for observer modeling.
• This AI-driven approach holds promise for improving the design and evaluation of medical imaging technologies like CBCT.