Deep learning for prediction of hepatocellular carcinoma recurrence after resection or liver transplantation: a discovery and validation study

Abstract

Background: There is a growing need for new improved classifiers of prognosis in hepatocellular carcinoma (HCC) patients to stratify them effectively.

Methods: A deep learning model was developed on a total of 1118 patients from 4 independent cohorts. A nucleus map set (n = 120) was used to train U-net to capture the nuclear architecture. The training set (n = 552) included HCC patients that had been treated by resection. The liver transplantation (LT) set (n = 144) contained patients with HCC that had been treated by LT. The train set and its nuclear architectural information extracted by U-net were used to train the MobileNet V2-based classifier (MobileNetV2_HCC_class). The classifier was then independently tested on the LT set and externally validated on the TCGA set (n = 302). The primary outcome was recurrence free survival (RFS).

Results: The MobileNetV2_HCC_class was a strong predictor of RFS in both LT set and TCGA set. The classifier provided a hazard ratio of 3.44 (95% CI 2.01-5.87, p < 0.001) for high risk versus low risk in the LT set, and 2.55 (95% CI 1.64-3.99, p < 0.001) when known prognostic factors, remarkable in univariable analyses on the same cohort, were adjusted. The MobileNetV2_HCC_class maintained a relatively higher discriminatory power [time-dependent accuracy and area under curve (AUC)] than other factors after LT or resection in the independent validation set (LT and TCGA set). Net reclassification improvement (NRI) analysis indicated MobileNetV2_HCC_class exhibited better net benefits for the Stage_AJCC beyond other independent factors. A pathological review demonstrated that tumoral areas with the highest recurrence predictability featured the following features: the presence of stroma, a high degree of cytological atypia, nuclear hyperchromasia, and a lack of immune cell infiltration.

Conclusion: A prognostic classifier for clinical purposes had been proposed based on the use of deep learning on histological slides from HCC patients. This classifier assists in refining the prognostic prediction of HCC patients and identifies patients who have been benefited from more intensive management.

Keywords: Deep learning; HCC; LT; MobileNetV2; Prognosis.

Fig. 1

The pipeline for MobileNetV2_HCC_Class. From the small image patches in 224 × 224 pixels of the train set, the heat map of nuclei segmentation for each tile was obtained using a pre-trained U-net. The heatmap of nuclei segmentation and the color-normalized RGB tiles were concatenated at channel level, and a four-channel tile was obtained. Bags containing four-channel tiles were then dumped into a feature extractor of the MobileNetV2 model. A generalized mean with a sign was used as the aggregation function

Fig. 2

Prognostic value of MobileNetV2_HCC_Class in LT set and the stratification of common prognostic variables. MobileNetV2_HCC_Class categorized patients into low-risk and high-risk subgroups. The prognostic value for MobileNetV2_HCC_Class was conservative, even following the stratification of common clinical and pathological variables. AFP: alpha-fetoprotein, Tumor No: tumor number, Diameter: total tumor diameter

Fig. 3

The performance of different risk factors for tumor recurrence after LT. The time-dependent accuracy (a) and AUC value (b) for different criteria based on tumor recurrence. NRI (c) according to different factors compared with the Stage_AJCC. Stage AJCC: the American Joint Committee on Cancer, AFP: serum alpha-fetoprotein, Tumor_No: tumor number, Total_diameter: total diameter of the tumor

Fig. 4

Prognosis of MobileNetV2_HCC_Class in TCGA set and the stratification of common baseline variables. MobileNetV2_HCC_Class predicts RFS while also following the stratification of other common baseline variables. Stage AJCC: the American Joint Committee on Cancer, AFP: serum alpha-fetoprotein, VI: vascular invasion

Fig. 5

Performance of different risk factors in tumor recurrence after resection. The time-dependent accuracy (a) and AUC value (b) for different criteria based on tumor recurrence. NRI (c) according to different factors compared with the Stage_AJCC. Stage AJCC: the American Joint Committee on Cancer, TIL: tumor-infiltrating lymphocyte

Fig. 6

Typical tiles were found to have low or high risks usingMobileNetV2_HCC_Class. Four hundred tiles with the highest predictability were investigated. The features used to predict high recurrence risk were stroma (a), cellular atypia (b), and nuclear hyperchromasia (c). The feature indicating low recurrence risk was the existence of immune cells (d)