Unifying spatial transcriptomics and H&E stain data under a common coordinate system for visual analysis and 3D reconstructions
Spatial transcriptomics (ST) is a powerful technology that allows researchers to visualize and quantify mRNA expression across a tissue. This technology helps researchers to better understand the development of complex diseases like pulmonary fibrosis, a chronic lung disease that culminates in the physical and genetic reconstruction of the lungs. H&E stain images play an important role in aiding in the visual analysis of ST data, but they are oftentimes in a different coordinate space than spatial data. Thus, in order to more efficiently interpret spatial data is it necessary to align H&E images, and other histological features, with spatial transcriptomics data. This project explores various image alignment methods and objectively compares them through calculated “goodness of fit” scores. Computer vision methods such as OpenCV’s landmark detection and transforms proved to be a powerful tool for image alignment, when the correct pre-processing steps are in place. While OpenCV performed well for aligning healthy lung samples, it still struggles to find landmarks within pulmonary fibrosis affect samples. To tackle this challenge a python package called effortless landmark detection (ELD) was used and tested. Using ELD a machine learning algorithm was trained to detect landmarks in fibrotic lung samples. Multiple pre-processing steps and ML training schemes were compared and contrasted, to determine the best method of image alignment. My ongoing efforts aim to refine these methods to achieve robust and accurate micron-scale alignment for many types of tissue samples.
