The Cancer Genome Atlas project possesses a massive repository of tumor tissue samples from a variety of patients and cancer types. Though it is a rich source of information, doing systematic analysis requires sorting, categorizing, and filtering of thousands of images for minute visual clues that lead to deeper knowledge. Moreover, the matter is complicated by the fact that different methods were used to create the images in the first place, leading to inconsistencies from the start.
Researchers at the Lawrence Berkeley National Lab developed a computer vision system that sifts through the image samples and identifies tumor subtypes, as well as their heterogeneity. Additionally, the system applies genomic data to identify molecular correlates of each subtype and uses historic clinical data to predict patient outcomes based on the cellular signatures of the tumors.
More from LBNL:
“Our goals are to identify morphometric and architectural traits that can be predictive of a therapy. We’d also like to learn about the molecular signatures that lead to architectural aberrations,” says Bahram Parvin of Berkeley Lab’s Life Sciences Division. The development of the core computational module and the pipeline were led by Hang Chang and Gerald Fontenay, respectively, in Parvin’s Lab in the Life Sciences Division.
The core computational module works by extracting each cell from an image, and then profiling properties of each cell such as size, shape, and organization. In this way, the telltale characteristics of a specific tumor subtype are gleaned from a large cohort of images.
As recently reported, the scientists validated their pipeline by applying it to 377 whole-slide images from 146 patients who have an aggressive brain cancer called Glioblastoma Multiforme. The pipeline identified several tumor subtypes based on a range of cellular profiles. It also determined whether each subtype is predictive of a patient’s response to alternative therapy. Although the pipeline was developed in a high-performance computer language, it is compute intensive and required extensive use of the cluster operated by Berkeley Lab’s IT Division.
The scientists also created an online repository for these images, which also includes images of low-grade glial and kidney renal carcinoma tumor sections. The website allows for Google-map style zooming and panning of the tissue sections. The scientists next hope to layer more information onto the images, in addition to cellular structure, to provide a broader representation of the tumors’ characteristics and interactions between different components of tumor histology.
Press release: New Computational Pipeline Analyzes Tumor Images, May Help Predict Response to Cancer Therapy