Development and Validation of Artificial Intelligence Algorithms in Digital Pathology

Division Of Basic Sciences - Nci

Investigators

Linked publications, trials & patents

Abstract

Continuing my second full year in appointment as a Stadtman Tenure Track Investigator at the National Cancer Institute, Center for Cancer Research, I have established my lab within the Molecular Imaging Branch, implementing/utilizing computational equipment purchased at the end FY24 and expanding current initiatives to biologically validate AI algorithms through the acquisition of reagents for spatial sequencing. I have two postdocatoral fellows and have an open recruitment for additional fellows at this time. I have made significant progress in all research studies during this FY: AIM 1: Development of biologically-driven histomorphological features of metastatic prostate cancer (mPC) on digital pathology. One major challenge in metastatic prostate cancer is the robust identification of relevant morphological features of tumor cells due to the anticipated heterogeneity of both morphological features and molecular diversity. It is accepted in the prostate cancer literature that lineage plasticity plays a key role in the development of therapeutic resistance. However, until now there has been no comprehensive study comparing the morphological features of advanced prostate cancer using computational approaches. ). We continue to advance the understanding of metastatic prostate cancer through novel unsupervised AI methods applied to multi-modal datasets. Integrated analysis of genomic alterations and digital pathology images reveals unique nuclear features of lineage plasticity and development of adverse phenotypes, such as neuroendocrine prostate cancer. Utilizing rapid autopsy data for discovery (Chen et al, AIME 2024), we have demonstrated novel features of neuroendocrine differentiation drive poor prognosis in several external validation sets. We hypothesize that distinct patho-biological features add independent prognostic and predictive information relevant for the management of patients with advanced metastatic prostate cancer. We have curated multi-institutional cohorts of >500 patients undergoing biopsy of metastatic disease from University of Washington, Cleveland Clinic, Weill Cornell, National Cancer Institute, MD Anderson, and West Coast Dream Team. We utilize biomarkers identified in Aim 1, new biomarkers identified from end-to-end training, and plan to evaluate multi-model models (clinical and major genomic alterations) to predict the risk of death in 3 years. These works we anticipate to publish within the first half of FY25. AIM 2: Integrated multi-scale biomarkers for outcome prediction and response assessment. We have published two articles to date for cancer detection, segmentation, and Gleason grading (PMID 38953042, PMID 39682085). We have now focused our attention on the development of nuclear grading and detection of lymphovascular invasion, which are independently prognostic factors in prostate cancer and such ultrastructural evaluations are time-consuming tasks for the anatomic pathologists. Ultimately, we would like to make these AI models available for clinical cancer care. We focus our current efforts on prediction of patients at highest risk for development of metastasis, validating in various external cohorts (NCI, UCLA, ACS, University of Washington, and two public cohorts [TCGA, LEOPARD]). We have identified that models trained from metastasis-free survival are more generalizable than those trained from recurrence-free survival endpoints. We now aim to validate our novel AI-based features identified in Aim 2 with underlying molecular profiling. Specifically, in a cohort of 30 patients we recently initiated a study performing matched spatial RNA sequencing (Visium V2) in distinct tumor regions to identify underlying molecular correlates of our AI predictions. While the architectural patterns identified by our ongoing research are expected to be distinct from those in the metastatic setting, there is a unique opportunity to compare signatures characterized in the metastatic cohort with those in localized disease. Digital scanning of patients with pre-treatment MRI undergoing surgical intervention at NCI are currently ongoing in collaboration with Dr. Baris Turkbey. AIM 3: Exploratory translation cancer for identification of multi-modal imaging. We have expanded our research efforts to localized bladder cancer, focusing on novel digital pathology features associated with response to neoadjuvant chemotherapy in transurethral resection/biopsy samples prior to treatment. This work identifies key features in a disease setting where there is currently a limited role for morphological features in patient risk stratification. Technically, we have made advancements in utility of novel AI frameworks (Zhang et al, PathVisions 2024 oral presentation). We showed that current foundation models do not translate well to bladder biopsies newly diagnosed patients, which have unique imaging appearance and artifacts due to resection technique. We have also demonstrated these signatures have underlying molecular correlates from TCGA validation studies and plan to evaluate these associations in larger cohorts (SWOG clinical trial) in the future. We have recently received approval for a protocol for development of AI algorithms in pan-cancer settings with the Joint Pathology Center (DoD), which currently holds >4 million slides. Initial models will be trained by WSIs of nearly 10,000 patients. These include many rare disease subtypes which are poorly represented in existing foundation models. We will evaluate the ability to fine-tune and develop novel AI models for development of diagnostic algorithms.