Modeling human observer performance with neural network observers in a forced localization task using undersampled
Sandro Amaglobeli1,2, Justine P Prasad1, Craig K Abbey3
1Mathematics Department, Hofstra University, Hempstead, NY, USA.
Summary
This study models human performance in magnetic resonance imaging (MRI) localization tasks. Training models on diverse undersampling conditions improved generalization and matched human performance, crucial for optimizing MRI acquisition.
Area of Science:
- Medical Imaging
- Computational Neuroscience
- Artificial Intelligence
Background:
- Accurately modeling human performance in forced localization (FL) tasks is challenging, especially with anatomical images and out-of-distribution data like undersampled magnetic resonance imaging (MRI).
- Generalizing AI models to varying levels of undersampling (0% to 20% low frequencies) in MRI is crucial for clinical applications.
Purpose of the Study:
- To develop and evaluate AI models that can accurately predict human performance in forced localization tasks using undersampled MRI.
- To investigate the impact of training data diversity (varying undersampling percentages) on model generalization and human-like performance.
Main Methods:
- Utilized a modified EfficientNet-B1 architecture (FLNet) and a biologically inspired V1Block preprocessing module (V1FLNet) for coordinate prediction.
- Trained and evaluated models on MRI images with controlled undersampling percentages (0% to 20% low frequencies).
- Compared model performance against average human performance across different training and testing conditions.
Main Results:
- Models trained on diverse undersampling conditions (0% and 20%) showed improved generalization compared to single-condition training.
- Training on all tested undersampling conditions yielded the best results, closely matching human observer performance.
- V1FLNet demonstrated comparable performance to FLNet, particularly when trained on all conditions.
Conclusions:
- Training AI models on a comprehensive set of undersampling conditions enhances their ability to generalize and mimic human performance in MRI localization tasks.
- The findings suggest that AI models trained on diverse data can inform optimal MRI acquisition strategies, mirroring human observer choices.
- Biologically inspired preprocessing (V1Block) did not significantly outperform standard architectures but showed similar efficacy when trained comprehensively.
