RT Journal Article
SR Electronic
T1 Using transfer learning on whole slide images to predict tumor mutational burden in bladder cancer patients
JF bioRxiv
FD Cold Spring Harbor Laboratory
SP 554527
DO 10.1101/554527
A1 Xu, Hongming
A1 Park, Sunho
A1 Lee, Sung Hak
A1 Hwang, Tae Hyun
YR 2019
UL http://biorxiv.org/content/early/2019/02/19/554527.abstract
AB The tumor mutational burden (TMB) is a genomic biomarker, which can help in identifying patients most likely to benefit from immunotherapy across a wide range of tumor types including bladder cancer. DNA sequencing, such as whole exome sequencing (WES) is typically used to determine the number of acquired mutations in the tumor. However, WES is expensive, time consuming and not applicable to all patients, and hence it is difficult to be incorporated into clinical practice. This study investigates the feasibility to predict bladder cancer patients TMB by using histological image features.We design an automated whole slide image analysis pipeline that predicts bladder cancer patient TMB via histological features extracted by using transfer learning on deep convolutional networks. The designed pipeline is evaluated to publicly available large histopathology image dataset for a cohort of 253 patients with bladder cancer obtained from The Cancer Genome Atlas (TCGA) project. Experimental results show that our technique provides over 73% classification accuracy, and an area under the receiver operating characteristic curve of 0.75 in distinguishing low and high TMB patients. In addition, it is found that the predicted low and high TMB patients have statistically different survivals, with the p value of 0.047. Our results suggest that bladder cancer patient TMB is predictable by using histological image features derived from digitized H&E slides. Our method is extensible to histopathology images of other organs for predicting patient clinical outcomes.