ML-based IDIF computation for dynamic FDG-PET of human brain; Privacy-Preserving Machine Learning in Biomedicine

Positron emission tomography (PET) is an imaging modality that can be used by clinicians to assess a variety of disease states in the brain, including epilepsy, glioblastoma and concussion. To conduct a new, more diagnostically useful form of brain PET, dynamic PET (dPET), the carotid arteries must be segmented out of the scan in order to allow for quantitative analysis of the physiological pathway of interest. Currently, these annotations are done manually, which is both a time-consuming and imprecise process, in turn slowing the clinical adoption of brain dPET and missing the opportunity for more accurate diagnoses. In order to facilitate the wider adoption of dPET, an image pre-processing pipeline and deep feedforward artificial neural network were created to automatically identify and segment these carotid regions of interest. With the decision to approach this problem with machine learning (ML) techniques, however, lack of diversity within the training data and poor model robustness can lead to inaccurate assessments of different disease states and even result in unnecessary medical intervention.

Privacy-preserving techniques (PPTs) offer the ability to combat the susceptibility of such models to similarly poor robustness. The four pillars of biomedical ethics, a framework intended to guide a clinician’s decision making process in their practice, will be applied to better understand the vulnerabilities and subsequent risks that arise when implementing various. In order to conduct this research, a literature review will be performed of technical papers explaining a number of PPTs. In particular, through analyzing the various benefits and vulnerabilities associated with each form of PPT being investigated, it is expected that similarities in the ways PPTs affect different stakeholders in healthcare, either positively or negatively, will become apparent. Such findings would certainly be useful for ML researchers working in the biomedical space, a perfect example being the previously discussed dPET segmentation pipeline being developed. For this project, the dataset available for model training is quite limited and could benefit from PPT’s ability to augment training dataset size. With this research, there could be better understanding amongst researchers of how the many forms of PPTs affect different stakeholders, helping guide their decision for which one to use (if any) and facilitate the selection of the most ideal PPT specific to their given project.

School of Engineering and Applied Science
Bachelor of Science in Biomedical Engineering
Technical Advisor: Bijoy Kundu
STS Advisor: Rider Foley