Development of a deep pathomics score for predicting hepatocellular carcinoma recurrence after liver transplantation

Tumor recurrence after liver transplantation (LT) impedes the curative chance for hepatocellular carcinoma (HCC) patients. This study aimed to develop a deep pathomics score (DPS) for predicting tumor recurrence after liver transplantation using deep learning.

Two datasets of 380 HCC patients who underwent LT were enrolled. Residual convolutional neural networks were used to identify six histological structures of HCC. The individual risk score of each structure and DPS were derived by a modified DeepSurv network. Cox regression analysis and Concordance index were used to evaluate the prognostic significance. The cellular exploration of prognostic immune biomarkers was performed by quantitative and spatial proximity analysis according to three panels of 7-color immunofluorescence.

The overall classification accuracy of HCC tissue was 97%. At the structural level, immune cells were the most significant tissue category for predicting post-LT recurrence (HR 1.907, 95% CI 1.490-2.440). The C-indices of DPS achieved 0.827 and 0.794 in the training and validation cohorts, respectively. Multivariate analysis for recurrence-free survival (RFS) showed that DPS (HR 4.795, 95% CI 3.017-7.619) was an independent risk factor. Patients in the high-risk subgroup had a shorter RFS, larger tumor diameter and a lower proportion of clear tumor borders. At the cellular level, a higher infiltration of intratumoral NK cells was negatively correlated with recurrence risk.

This study established an effective DPS. Immune cells were the most significant histological structure related to HCC recurrence. DPS performed well in post-LT recurrence prediction and the identification of clinicopathological features.