Quantitative Imaging in Pathology: Mapping and Characterizing Spatial Patterns in Tissue Images for Cancer Research

Stony Brook University: Joel Saltz, Tahsin Kurc, Raj Gupta, Erich Bremer, Feiqiao Wang, Tammy DiPrima, Joseph Balsamo, Dimitris Samaras, Le Hou, Han Le, Vu Nyugen, Shahira Abousamra, Maozheng, Zhao

Emory University: Ashish Sharma, Ryan Birmingham, Nan Li, Whitney Hogg

Oak Ridge National Laboratory: Scott Klasky, Jeremy Logan

Digital Pathology is a rapidly growing field with applications in translational and clinical research. The automated quantification of pathology phenotype is a crucial component of Digital Pathology towards realizing the goal of precision medicine. Our team has developed analysis pipelines and a suite of supporting tools in an integrated software framework, called QuIP, in order to extract and characterize image-based phenotypes from high resolution whole slide tissue images. Our analysis pipelines include convolutional neural networks (CNN): (1) for characterization of tumor regions; (2) segmentation of nuclear material and nuclei; (3) detection and characterization of spatial lymphocyte patterns; and (4) characterization of cellular patterns in images of multiplexed Immunohistochemistry stained tissue specimens. These core analysis methods are augmented by a variety of methods to compute summary spatial statistics at patch, image and patient levels for correlation with clinical and genomic data. All of our analysis pipelines are implemented as containers for easy deployment and execution on local machines and in the Cloud. We also have developed QuIP, an integrated software infrastructure consisting of web applications and fully containerized set of micro-services. QuIP provides capabilities to (1) manage images and metadata about images and image analysis products (PathDB), (2) manage analysis results (FeatureDB), (3) interactively visualize and annotate high-resolution tissue images (caMicroscope), and (4) visualize and interact with spatial patterns of lymphocytes and tumor regions (FeatureMap). Collectively QuIP services allow users to interact with large volumes of image and analysis results data, review analysis output for quality, and generate training data to train or refine deep learning models. The components of the QuIP infrastructure and analysis pipelines are being integrated in the next generation software stack, called PRISM, for The Cancer Imaging Archive (TCIA). These new capabilities will facilitate integrated analyses and use of Radiology and Digital Pathology data in cancer research. Combined use information from Radiology and Digital Pathology can enable a more robust multi-scale characterization of cancer physiology and new multi-scale insights into cancer processes.

Source codes for the analysis pipelines and QuIP software infrastructure can be accessed at the following link.

QuIP Software Distribution: https://github.com/SBU-BMI/quip_distro
Analysis Pipelines: https://github.com/SBU-BMI/quip_analysis

This work is supported in part by 1U24CA180924-01A1, 3U24CA215109-02, and 1UG3CA225021-01 from the National Cancer Institute, NCIP/Leidos 14X138, and R01LM011119-01 and R01LM009239 from the U.S. National Library of Medicine.