An ensemble deep learning model for tumor bud detection and risk stratification in colorectal carcinoma.

点击翻译按钮获取中文摘要

PROJECT ABSTRACT Colorectal cancer (CRC) is the fourth most common cancer, and the second leading cause of cancer death in the United States, with an estimated incidence of 151,030 new cases in 2022. According to the American Cancer Society, the lifetime risk of developing colorectal cancer is 1 in 23 for men and 1 in 25 for women. Tumor budding is a prognostic factor in colorectal cancer with potential to risk stratify patients and possibly guide treatment decisions. It is defined as the presence of a single tumor cell or a cell cluster consisting of fewer than five tumor cells at the invasive tumor front. Unfortunately, tumor budding is not routinely disclosed in pathology reports due to lack of reproducible methods in identifying tumor buds from H&E slides. The prevalence, mortality, and risk of colorectal cancer as well as the potential of tumor budding as a prognostic factor necessitate an accurate, easy- to-use, reproducible system to identify tumor budding. We aim to develop a computer-aided image analysis system to standardize the quantitative criteria used to define tumor budding from H&E slides. In addition to identifying tumor buds, the system will correlate tumor buds with several outcomes (microsatellite instability status, overall survival, progression free survival, and disease free survival). As part of the proposed computer- aided image analysis system, we will first develop a sophisticated method for color deconvolution to compensate for color variations. This will be followed by deformable image registration and deep learning modules to differentiate tumor from non-tumor regions. The study will show that machines can be trained using deep learning to identify different anatomical regions within H&E slides of colorectal patients. From thereon, we will rely on scale-space theory and alpha-shapes to identify tumor buds and hotspots. We will use mathematical morphology and differential geometry to extract visually meaningful imaging features from tumor buds and hotspots. We will explore the potential of these imaging features along with features produced by our unsupervised multiple instance learning in predicting outcomes. The proposed research will help identify the association of tumor budding to colorectal cancer outcomes. The model will be subjected to rigorous statistical analysis for accuracy and reproducibility. The project will result in innovative software tools that facilitate the selection for personalized cancer therapies for colorectal patients.