Do Sites Benefit Equally from Distributed Learning in Medical Image Analysis?

Do Sites Benefit Equally from Distributed Learning in Medical Image Analysis?

연구 분야: Databases

학회: MICCAI Workshop on Fairness of AI in Medical Imaging, Workshop on the Ethical and Philosophical Issues in Medical Imaging

Artificial intelligence (AI) has the potential to make medical image analysis more accessible to healthcare institutions worldwide. However, when trained on multi-site datasets, models may excel with data from certain institutions but struggle with others, even when exposed to their training data. This emphasizes the importance of investigating whether all sites benefit from AI models, especially within distributed learning setups. Distributed learning methods allow access to broader and more diverse datasets from multiple sites during training. In this context, the travelling model (TM) paradigm has demonstrated superior performance in limited data scenarios, making it particularly relevant in low-resource settings. This work investigates whether all sites can benefit from TM development and identifies the factors influencing performance disparities. Specifically, a Parkinson’s disease (PD) database comprising 1,817 neuroimaging datasets from 83 different sites is utilized to investigate how site-specific and participant-specific factors influence the performance of TM in classifying PD. Therefore, we analyze the false positive rate (FPR) and false negative rate (FNR) to identify the characteristics contributing to misdiagnosis. Our findings reveal disparities in benefits across sites, with class imbalance emerging as the major factor influencing FPR and FNR, especially for sites with more PD cases. This research underscores the urgency of a rigorous analysis of a model’s behaviour in distributed setups to detect misdiagnosis risks and encourage developers to evaluate and optimize models beyond overall accuracy. Thus, comprehensive analyses of this type can help pave the way for the development of more equitable AI-based medical imaging models.