ABSTRACT
Label-free optical microscopy has emerged as a promising method for rapid imaging of fresh, unprocessed
surgical specimens. Stimulated Raman histology (SRH) – a label-free, non-destructive, high-sensitivity optical
imaging method – combined with artificial intelligence (AI) has been used for bedside brain tumor diagnosis,
margin delineation, and molecular genetic prediction. Previous AI methods are limited because they rely on
weak, slide- or patient-level annotations for model training. Importantly, these annotations fail to capture the
cellular complexity and spatial heterogeneity found in diffuse gliomas, the most common and deadly primary
brain tumor. A major barrier to advancing the role of label-free optical imaging in diffuse glioma research and
patient treatment is developing strategies to allow AI-based computer vision models to learn rich, high-
resolution, single-cell optical image features, which allow for a more complete description of the underlying
tumor biology. The objective of this research is to determine if an AI-based computer vision system can learn
single-cell optical image features. I will (1) detect single cells in diffuse glioma specimens imaged using SRH;
(2) optimize learned single-cell features using representative single-cell examples, which we call exemplar
learning; and (3) identify patient-level clusters based on optical single-cell features. Successful completion of
this proposal will reduce the reliance of AI on weak annotations, and advance the role of AI in diffuse glioma
research and treatment via single-cell optical features and patient-level clusters. In the long term, AI methods
in this proposal can be integrated with the diagnosis workflow using optical imaging, and provide physicians
with additional tools to further differentiate diffuse gliomas, and enable better and more personalized care.
