82MO - Automatical risk stratifying for colorectal cancer by deep learning based pathological score

Background

Histology whole-slide images contains crucial information of patients' risk. In this study, we aimed to develop a fully automated approach for stratifying colorectal cancer (CRC) patients’ risk of mortality by applying deep learning convolutional neural networks to histology whole-slide images.

Methods

We developed deep learning model of two-stage for stratifying CRC's survival risk. First a tumor differentiation classifier was trained on expert annotations to generate probability maps and a predictor index for tumor differentiation. Survival times and aforementioned maps were used to train a Cox neural network that predicts overall survival. Three CRC datasets were used including 2 public datasets from The Cancer Genome Atlas (TCGA); TCGA-COAD was used for model training, TCGA-READ and a local institutional dataset were used for validation.

Results

The accuracy of predicting tumor differentiation achieved to 85% (AUCs 0.983, 0.963, 0.963 and 0.981 for low, medium, high grades, and normal tissue, respectively) in validation set of TCGA-COAD. The patient specific score that predicted by Cox survival CNN networks showed it;s moderate discrimination ability with c-index of 0.65 in validation set. The score can also be combined with clinical factors (age, gender and TNM staging) as a powerful prediction model to stratify CCR's risk. The combined Cox model archieved c-indices of 0.77 and 0.79 in two independent validation sets. C-indices of the model in two validation sets were 0.72 and 0.75 without deep learning based pathological score.

Conclusions

In this study, we proved that the deep learning based psthological scores were added factor beyongd normal clinical factors of sex, age and TNM staging and significantly enhanced the prediction performance of these factors. This study demonstrates that deep learning can provide a meaningful reference for tumor differentiation grading and patient risk evaluation for pathologists.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Fudan University Shanghai Cancer Center.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.