A deep learning anthropomorphic model observer for a detection task in PET.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Lesion detection in Positron Emission Tomography (PET) is crucial for oncology.
• Anthropomorphic model observers (MOs) assess task-based image quality by mimicking human observers (HOs).
• Deep learning MOs (DLMOs), particularly CNNs, are emerging for various imaging modalities, but their use in PET is less explored.

Purpose of the Study:

• To evaluate if DLMOs can better predict human observer performance than conventional MOs in PET lesion detection.
• To assess DLMOs in a two-alternative forced-choice (2AFC) detection task using PET images with realistic anatomical variations.

Main Methods:

• Two DLMOs were developed: a CNN-based DLMO and a CNN-SwinT DLMO integrating CNN and Swin Transformer encoders.
• PET images with and without simulated lesions were used, with labels provided by eight human observers (radiologists and image scientists).
• Performance was compared against Channelized Hotelling Observer (CHO) and Non-Prewhitening Matched Filter (NPWMF) using metrics like prediction accuracy and Mean Squared Error (MSE) in a 9-fold cross-validation.

Main Results:

• Both CNN DLMO and CNN-SwinT DLMO outperformed CHO and NPWMF in accuracy and MSE.
• The CNN-SwinT DLMO demonstrated the highest prediction performance among all evaluated model observers.

Conclusions:

• DLMOs offer superior prediction of human observer performance in PET lesion detection compared to traditional MOs like CHO.
• Integrating Swin Transformer with CNN encoders significantly improves DLMO prediction accuracy over CNN-only approaches.