Digital score of lymphocytic infiltration in tumor-associated stroma in relation to overall survival in bladder cancer.

Abstract

4576 Background: Bladder Urothelial Carcinoma (BLCA) is the most common type of bladder cancer and the sixth most frequent cancer in the US. High numbers of lymphocytes colocated with tumour-associated stromal (TAS) tissue has demonstrated prognostic significance for overall survival in multiple cancers. Our goal in this study was to explore the prognostic significance of an automated score to quantify the colocalisation of lymphocytes and TAS for overall survival (OS) in BLCA patients from Haematoxylin & Eosin (H&E) stained Whole Slide Images (WSIs). Methods: Two cohorts of BLCA patients were included in this study. From the UK, a cohort of 67 BLCA patients constituted cohort A. We also evaluated our method on The Cancer Genome Atlas (TCGA) bladder cancer cohort of 453 cases, which we refer to as cohort B. We developed a two-stage method for digital quantification of lymphocytic infiltrates in TAS. First, we employed an AI algorithm to recognise and classify different regions as areas of high concentration of tumour, lymphocytes and stroma for each WSI creating a segmentation map of the different tissue types. For patients with multiple WSIs, we used the slide with the highest percentage of tumour to predict survival. This algorithm had been pre-trained on a cohort of Oral cancer and was fine-tuned using annotations from 4 BLCA WSIs, 3 from cohort A and 1 from cohort B. These WSIs were excluded from the final survival analysis. Using the segmentation maps, a statistical measure for the colocalisation of TAS and lymphocytes termed the tumour-associated stroma infiltrating lymphocytes or (TASIL) score was computed. Finally, for each cohort, data was right-censored at 10 years and the digital BLCA-TASIL (BT) score’s prognostic significance for OS was investigated by fitting a Kaplan-Meier estimator and Cox proportional hazard (PH) analysis, stratifying patients into two groups based on the BT score. In each cohort, two thirds of the data was used as the discovery set to determine the best cut-off for the TASIL Score and the remaining third was used as the validation set. Results: Our classification algorithm achieved high average F1-score of 0.88 on a held-out set of unseen data for the classification of tumour, lymphocytic and stromal regions. In cohort A, higher BT score was strongly associated with better OS ( P= 0.00906) on the unseen validation data. This significant association was also found in the validation data of cohort B ( P< 0.001). Using the BT score as the only covariate, a Cox PH model for the validation data resulted in a C-index of 0.73 for cohort A and 0.57 for cohort B, respectively. Conclusions: The digital BT score showed significant prognostic value for overall survival in both BLCA cohorts, reinforcing the findings of prior work. We intend to further validate these findings on another cohort. To the best of our knowledge, this is the first attempt to predict survival solely from H&E slides in BLCA.