P-L12 Exploring Pathological Signatures for Predicting Recurrence of Early-stage Hepatocellular Carcinoma Based on Deep Learning

Abstract

Abstract Background Early-stage hepatocellular carcinoma (HCC) is the ideal indication for liver resection. High recurrence rate limits the radical possibility. Current clinicopathological determinants are insufficient to reliably evaluate the recurrence risk after surgery. To address this global issue, we aimed to use deep learning to explore novel pathological signatures based on histological slides for predicting early-stage HCC recurrence and to evaluate the relationship between histological features and multi-omics information. Methods 576 pathological images collected from 547 patients with BCLC stage 0-A HCC who underwent hepatectomy from 2006 to 2015 were randomly divided into the training cohort and validation cohort. The external validation cohort was composed of 147 TNM I patients from TCGA database. Weakly supervised convolutional neural networks were used to identify six classes of HCC tissues. Pathological signatures were extracted and two novel risk scores were constructed by LASSO Cox to predict recurrence. The forecast performance of the scores and patients' prognosis were evaluated. The relationship between histological score (HS) and immune infiltrating cells was estimated by clustering analysis. Results The classification accuracy of HCC tissue was 94.17%. The C-indexes of histological score in the training, validation and TCGA cohorts were 0.804, 0.739 and 0.708, respectively. Multivariate analysis showed that microvascular invasion (HR = 1.46, 95% CI: 1.09-1.95) and HS (HR = 4.05, 95% CI: 3.40-4.84) were independent risk factors for recurrence-free survival. Patients in HS high-risk group had elevated alpha fetoprotein, worse tumor differentiation and higher proportion of microvascular invasion. HS was positively correlated with the expression of CD14 in adjacent normal liver tissue (P = 0.013), and negatively correlated with the expression of CD8 in tumor (P < 0.001). Conclusions This study established and validated two novel risk scores based on the histological slides using deep learning. HS performed well in recurrence prediction for early-stage HCC patients and indication of important clinicopathological features.