Reducing Bias in AI Algorithms for Gallium-68 PET: A Bioethical Perspective Using Transfer Learning

Abstract This is a supplemental application under NOT-OD-25-015 to conduct bioethics research and advance capacity building related to bias in AI algorithms. This project is specific to both bioethics research and capacity building in bioethics. Bias in AI algorithms can lead to inaccurate predictions, delaying diagnoses, misguiding treatments, and worsening patient outcomes. Furthermore, biases in AI algorithms can disproportionately affect certain demographic groups, amplifying existing healthcare disparities and leading to inequitable outcomes. How to reduce bias, particularly due to insufficient or unrepresentative datasets, is a major concern in healthcare. Transfer learning offers a promising solution to mitigate data bias when datasets are small or unrepresentative. This is especially relevant for PET imaging, such as Gallium 68 PET scans, for which datasets are limited. With support from the parent R01 project, the team has developed a novel transfer learning framework for PET image quality enhancement. The proposed framework aims to perform pre-training using large-scale high-quality 18F- FDG PET datasets and then fine-tune on limited Gallium 68 datasets. In this supplement project, we will comprehensively evaluate the bias-reduction effect of this proposed framework, explore additional bias mitigation strategies, and investigate potential biases that may rise from transfer learning. Furthermore, we will actively disseminate our findings to research, educational and clinical communities. Bias in AI algorithms is a pressing and emerging bioethical issue when adopting AI to clinics. In the era of rapid AI advancements, this study will provide a robust evidence base to inform and guide future bioethical policies for AI/ML practices, particularly as pre-training and transfer learning become foundational in developing large healthcare AI models. Furthermore, this study will also advance AI-bioethics capacity building by: developing transferable frameworks and methodologies applicable to various biomedical applications, and creating educational resources to address bioethical challenges stemming from biases in AI algorithms.