Integration of epidemiology, pathology, immunology and outcomes in colorectal cancer

ABSTRACT Machine learning has the potential to transform pathologic diagnosis and to address very limited accessibility of expert pathology in low-income countries. Routine histology images of solid tumors contain an immense number of visual features that can be extracted and processed by artificial intelligence tools like machine learning, which excels at basic image analysis tasks such as tumor detection. In addition, machine learning can also predict clinically relevant features directly from histology images including microsatellite instability and immune features that independently predict prognosis response to therapy. This large, multicultural, racially and ethnically diverse study uses images of whole slides from routinely collected clinical specimens and applies computational pathology methods and digital spatial expression profiling to quantifiably improve CRC diagnosis, prognosis and predictive models together with clinical, epidemiologic and genetic data. The study goals will be accomplished through three specific aims. In Aim 1, we will apply novel machine learning algorithms from whole slide images to reproducibly identify MSI, histopathologic and immune features of colorectal cancer in racially/ethnically diverse populations. We will study H&E slides from 6,751 CRC cases, digitizing existing slides from 5,551 CRC cases and 1,200 new cases of CRC with contemporaneous clinical and epidemiologic data. Then, we will apply deep learning methods to accurately identify histopathologic features and immune characteristics of CRC. We will use a robust training validation, and testing design (70%/15%/15%) to ensure the rigor and reproducibility of our findings. In Aim 2, we will test whether machine learning algorithms that predict MSI and immune features related to CRC prognosis improve with the addition of clinical, epidemiologic, and germline genetic data. We will use machine learning statistical methods to test whether algorithms developed in Aim 1 improve prediction of overall survival and response to therapy with the addition of supplemental information beyond whole slide digital images. Finally, in Aim 3, we will compare the information derived from digital spatial profiling of expressed proteins in colorectal tumors with the information derived from Immunoscore quantification of lymphocyte populations at the tumor center (CT) and the invasive margin (IM), and explore whether these measures improve the models developed in Aims 1 and 2 in a subset of samples. We will perform GeoMx digital spatial profiling of 56 proteins expressed in 150 Stage I-III TNM colorectal cancers to compare the performance of digital spatial profiling to Immunoscore, a scoring system relying exclusively on expression patterns of CD3+ and CD8+ T cells. This study takes advantage of pathologic, epidemiologic, clinical, immunologic and germline genetic data from racially/ethnically diverse CRC patients from California, Detroit, New York, Florida, Puerto Rico, Israel and Spain. Our overarching goal is to improve the efficient diagnosis of colorectal cancer with clinically impactful immune profiles.