Leveraging artificial intelligence/machine learning-based technology to overcome specialized training and technology barriers for the diagnosis and prognostication of colorectal cancer in Africa

点击翻译按钮获取中文摘要

SUMMARY/ABSTRACT Colorectal cancer (CRC) is the third most commonly diagnosed cancer and the second leading cause of cancer related deaths worldwide. Rates in Africa are on the rise, but essential histopathology services critical for cancer care are scarce. To address this barrier, we developed an artificial intelligence (AI)/machine learning (ML)-based computational pipeline (SIVQ/VIPR) that performs automated pixel-level image segmentation and classification from digital images of routinely collected hematoxylin and eosin (H&E)-stained slides. SIVQ/VIPR is highly precise, reproducible, and outperforms subject matter experts. Once histologically distinct regions are identified, image analysis algorithms can then identify individual regions and aggregate them to predict diagnostic and prognostic features in conjunction with clinical outcomes to guide treatment. Our overall approach is to leverage our validated SIVQ/VIPR computational pipeline to develop and validate an AI-based diagnostic decision support (AI-DDS) tool for CRC diagnosis and prognosis in an existing Kenyan cohort. To carry out this work, the Aga Khan University (AKU)- East Africa and the University of Michigan have partnered with Tenwek Hospital, a non-academic community-based public hospital in rural Bomet, Kenya, to develop a unique collaboration of oncologists, pathologists, surgeons, statisticians, and informaticians, making us uniquely suited to develop population-relevant, affordable, and scalable data science solutions in Kenya – all priorities of the DS-I Africa Program. We will: Aim 1. Adapt and validate an existing ML-based diagnostic algorithm for CRC using digital fields of view from H&E-stained slides in a retrospective cohort of n=675 CRC cases from the AKU and Tenwek Hospitals. We will apply the CRC-trained SIVQ/VIPR computational pipeline for segmentation and classification for CRC features, followed by a confirmatory classifier step to achieve a case level, binary result of a cancer/no-cancer (i.e., diagnosis). Aim 2. Develop and refine an unsupervised ML method to identify histopathology image-derived measurements associated with CRC prognosis. We will use computer/machine vision approaches to identify image features (e.g., cellular morphology) discriminative of CRC prognosis and biological potential for disease aggressiveness. Combined use of AI-based morphological signatures of aggressive disease (e.g., high-grade tumor architecture) will be compiled with other clinically relevant features towards the goal of generating a multi-axial multiplexed AI-DDS tool that can maximally inform of the biological and metastatic potential of each CRC case. This project will lay the groundwork for an AI-DDS tool for clinicians (e.g., pathologists, oncologists) that facilitates prompt and accurate diagnosis, prognosis, and risk stratification for CRC care in Africa. Because this approach leverages open-source software and can be deployed as a turn-key system intended for web-based cloud deployment, it is well-suited for capacity building, integrating into educational programs, and expanding to other emergent or prevalent cancers (i.e., breast, cervical, prostate) as part of the DS-I Africa Consortium.