Photoacoustic device fingerprints induce bias in deep learning models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning (DL) models excel in established medical imaging but struggle with emerging modalities like photoacoustic imaging (PAI) due to data sparsity and bias.
• Hardware-induced bias in PAI is an unexplored challenge, risking multicentric deployment failure.

Purpose of the Study:

• To conduct the first multicentric analysis of hardware-induced bias in photoacoustic imaging.
• To investigate how device-specific characteristics affect DL model performance in PAI.
• To evaluate the impact of device-health correlations on diagnostic accuracy.

Main Methods:

• Analyzed device-specific characteristics from four PAI device instances across two peripheral artery disease studies.
• Trained DL models to classify device origin and diagnose disease under varying device-health correlations.
• Assessed the ability of DL models to detect device-specific fingerprints.

Main Results:

• Identified unique 'fingerprints' embedded in PAI images from different device instances.
• Achieved high accuracy in DL-based device detection using these fingerprints.
• Demonstrated that DL models exploit device signatures as shortcuts when device and health status are correlated, leading to biased predictions.

Conclusions:

• Overlooking hardware-induced bias in PAI can lead to overestimated algorithm performance and clinically misleading results.
• Emphasizes the critical need for bias evaluation and explainable AI methods in PAI research.
• Facilitates the development of robust and reliable DL models for multicentric photoacoustic imaging studies.