Deep Learning Model Observers Trained with Human Observer Data from Two-Alternative Forced Choice (2AFC) Trials

• Published in: 2023 IEEE Nuclear Science Symposium, Medical Imaging Conference and International Symposium on Room-Temperature Semiconductor Detectors (NSS MIC RTSD) p. 1
• Main Authors: Shao, M., Byrd, D. W., Abbey, C. K., Mitra, J., Wollenweber, S. D., Behnia, F., Lee, J. H., Iravani, A., Sadic, M., Chen, D. L., Kinahan, P. E., Ahn, S.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.11.2023
• Subjects: Deep learning; Neural networks; Observers; Receivers; Semiconductor detectors; Semiconductor device modeling; Training
• ISSN: 2577-0829
• DOI: 10.1109/NSSMICRTSD49126.2023.10338151

A model observer, designed to predict human observer responses in a detection task, is an efficient tool to characterize the task-based image quality, and is also useful for designing a system, a protocol, and an image reconstruction algorithm to optimize the detection performance. Conventionally, linear model observers such as channelized Hotelling observer (CHO) have been widely used. Recently, as a new type of model observers, neural networks have been trained to predict human responses in a detection task. Often, two-alternative forced choice (2AFC) experiments are performed, which have a useful property that the percent correct of 2AFC trials is equivalent to the area under the receiver operating characteristic curve under mild conditions. However, it is not straightforward to optimally train a neural network using 2AFC human observer data. In this study, we present a training scheme for a deep learning model observer (DLMO) directly using 2AFC data, where two neural networks with shared weights are trained. We evaluate the training scheme using 2AFC data from 8 human observers including 4 radiologists with PET images generated using simulated lesions with clinical backgrounds. This study shows the DLMO trained directly using the 2AFC data can predict human observer responses more accurately than the other model observers including CHO and a DLMO trained indirectly using the 2AFC data.