Abstract PO-001: Deep learning-based image analysis of the histological Glasgow Microenvironment Score in patients with colorectal cancer

Abstract

Abstract Background: Efforts to prognostically stratify colorectal cancer (CRC) patients using histological samples has led to the development of the Glasgow Microenvironment Score (GMS). However, the subjective nature of the GMS assessment criteria makes reproducing the score challenging. Therefore, the aim of the current study was to assess the clinical utility of an image analysis-based approach. Methods: In two cohorts of patients (a test cohort of 297 Stage II/III CRC patients from Norway and a validation cohort of 233 Stage II/III CRC patients from Scotland), one H&E section representing the deepest point of tumour invasion was assessed for Tumour Stroma Percentage (TSP) and the Klintrup-Makinen (KM) local immune grade to produce the GMS. A decision forest-based classifier was built to segment the tumour and stroma compartments, having been trained on a hand annotated subset of the test cohort. A UNET convolutional neural network was trained on the same training set as the decision forest algorithm to segment and quantify the tumour and stroma components. Each section was also annotated for the invasive margin and a manually built, threshold-based lymphocyte detection algorithm applied to quantify the local immune grade. Results: The decision forest classifier was able to stratify the test cohort patients for cancer specific survival (P = 0.009) based on the patients TSP, however failed to stratify patients from the validation cohort (P = 0.055) for TSP without retraining, therefore the deep learning approach was used for this cohort. Individually, the TSP (P = 0.001) generated from the UNET based classifier and the quantification of the local immune grade (P = 0.005), both significantly stratified validation cohort patients for survival. When the individual scores were combined into a digital version of the GMS, not only did the image analysis approach significantly stratify patient survival (P < 0.0001), it also significantly correlated with the manual scores both statistically (Chi Square = 32, P = 1.5 × 10−8) and in terms of survival estimates per patient group (GMS 0 - manual = 159 months vs digital = 159 months; GMS 1 – manual = 136 months vs digital = 139 months ; GMS2 – manual 101 months vs 97 months). To negate the need for any annotation prior to analysis, the lymphocyte detection algorithm was applied to the stroma areas determined by the UNET algorithm. This fully automated digital version of the GMS was able to stratify validation cohort patients into the 3 prognostic groups based on CSS (P < 0.001), however, GMS 0 had a lower survival estimate (automated = 148 vs annotated = 159 months). Conclusion: An H&E based image analysis approach to the GMS, on a single H&E slide, holds prognostic significance for CRC patients and correlates significantly with manual pathological assessment, making the assessment of the GMS more time efficient, reproducible, and clinically relevant. Citation Format: Christopher J. Bigley, John Waller, Alison Bigley, Karin Oien, Joanne Edwards, Antonia Roseweir. Deep learning-based image analysis of the histological Glasgow Microenvironment Score in patients with colorectal cancer [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-001.