BCCMA: VA Colorectal Cancer Clinical & Computational Collaborative (VA-5C): Project #2 - Artificial Intelligence and Multispectral Imaging of Colorectal Polyps.

Va Boston Health Care System, Boston MA

Investigators

Abstract

Objectives: This BLRD-CSRD-Collaborative MERIT Award (BCCMA) application is comprised of three projects unified by their focus on CRC precursors in Veterans. We, together with Los Angeles and San Diego VA’s call our team-science consortium the “VA5C,” for VA Colorectal Cancer Clinical & Computational Collaborative. Our multidisciplinary collaborative is linked by a common focus on CRC precursors as targets for intervention. The overarching objective of this proposal is to deploy and validate clinically an artificial intelligence (AI)-based low-cost platform to make endoscopic prevention of colorectal cancer (CRC) more efficient. We seek to create an accurate and widely adoptable, real-time histology (RTH) platform based on AI. Colonoscopic CRC prevention hinges on the complete removal and histopathological assessment of all polyps. This practice results in the removal of large numbers of polyps that have negligible malignant potential creating a need for tools to assess histology in real time to decrease biopsy costs and risks. Professional societies now endorse the purely optical management of small polyps and have put forth guidelines and performance thresholds toward adoption. Application of AI to RTH should help to make operators more consistent and accurate. For VA, such capability would finally open the door to widespread adoption of cost- saving resect-and- discard and leave-behind paradigms for diminutive polyp management. Our work on a Computer Aided Diagnosis (CADx) system that we call “EndoVet-AI” shows great potential to be a widely adoptable, accurate RTH platform in VA for distinguishing hyperplastic from adenomatous polyps. Sessile serrated lesions (SSLs), which are a distinct, visually subtle precancerous type of polyp, have not been addressed directly. To this point, we hypothesize that AI trained on SSLs will be capable of distinguishing their borders with reasonable accuracy. With this backdrop, the broad aims of Project #2 are to (Aim 1) Grow and maintain a CRC image and video database toward a formal curated VA-wide Image Repository for CRC-related AI research. (Aim 2) Collaborate with machine learning experts to deploy and validate clinical CADx models that differentiate benign serrated lesions (SLs) (e.g. hyperplastic polyps) from SSLs. (Aim 3) Pursue preclinical translational research with San Diego to define the typically subtle borders of SSLs in genetically engineered mouse models (GEMMs) using a multimodal mouse colonoscope and fluorescence imaging of mucin antibodies. The ultimate goal is to translate work in GEMMs combined with annotations of human SLs to develop AI-assisted segmentation to guide complete resection of detected SSLs. Methodology: We will (1) Create a Gastroenterology (GI) Image Data Repository which can house still images, videos of complete colonoscopies, scanned digital pathology slides, and associated procedure data; (2) initiate a multisite clinical study to acquire videos from patients at increased risk of SSLs, then use that data to design and validate AI algorithms for RTH of polyps to include SSLs; (3) develop preclinical translational AI models to delineate the indistinct borders of serrated lesions in GEMMs; and (4) prospectively test new algorithms with the two other VA facilities in the BCCMA. The primary endpoint of this last aim will be to evaluate the performance of CADx benchmarked against performance thresholds per professional society guidelines.