Discrete Frequency Infrared Spectroscopic Imaging for Breast Histopathology

PROJECT ABSTRACT Infrared (IR) spectroscopic imaging directly measures the chemical composition of cells and tissues for each pixel in the image. Using machine learning, this chemical data can be converted to pathology knowledge, without the use of dyes or stains – providing a potentially new avenue for clinical diagnoses and research to broadly aid public health. Since machine learning is integral to the approach, cognition of disease features can make diagnoses faster, cheaper and more precise. Interestingly, the approach can measure the tumor’s molecular characteristics and the microenvironment together in one shot. These capabilities can extend state of the art pathology practice by providing multiplexed stain-free molecular data and predictive models involving spatial and chemical information from multiple cell types. However, there are significant challenges and engineering development needed before this vision can be realized, including: (a) an imaging system that is competitive in measurement time with current clinical practice, (b) accurate and assured results that extend our ability beyond routine pathology, and (c) demonstration of robust use by pathologists and non-experts in technology. In the last project period(reported in 25 peer-reviewed publications, 2 granted patents), we developed “high-definition” (HD) IR imaging, which is now the standard commercial configuration for IR imaging manufacturers. We also developed the concept of “stainless staining” in which “low-definition” IR images appear to look like low-resolution stained images. We also demonstrated highly accurate breast tissue classification for a small number of pathologies. In this project period, we propose an advanced IR imaging system (newly designed optics, scanning) to make the technology powerful enough to provide a sample-to-image time of ~10 min for large surgical resections. This allows HD imaging in real time and will allow images, such as from stainless stains, be near the quality of those used by clinicians and researchers. Technological innovations lie in a design that is the first novel re-design of IR imaging in over 40 years and performance that is higher in speed, accuracy and image quality than ever before. Another critical part of our approach is to develop appropriate computational pipelinesfor extant problems in breast pathology. In addition to traditional models, we will validate the emerging tools of deep learning when appropriate. Finally, these technological realizations are followed by validation for a set of important problems in breast cancer care and research. The solutions will be rigorously evaluated against pathologist diagnoses, using high-quality, annotated data from 400 patients’ surgical resections and multiple tissue microarrays. Consequently, protocols for a number of identified pain points in breast pathology will result in addition to the technological progress, making the approach ready for use.