TED: Two-stage expert-guided interpretable diagnosis framework for microvascular invasion in hepatocellular carcinoma

Microvascular invasion (MVI) has been clinically recognized as a prognostic factor for hepatocellular carcinoma (HCC) after surgical treatment. Detection of MVI before surgical operation greatly benefit patients' prognosis and survival. Most of the existing methods for automatic diagnosis of MVI directly use deep neural networks to make predictions, which do not take into account clinical knowledge and lack of interpretability. To simulate the radiologists' decision process, this paper proposes a Two-stage Expert-guided Diagnosis (TED) framework for MVI in HCC. Specifically, the first stage aims to predict key imaging attributes for MVI diagnosis, and the second stage leverages these predictions as a form of attention as well as soft supervision through a variant of triplet loss, to guide the fitting of the MVI diagnosis network. The attention and soft supervision are expected to jointly guide the network to learn more semantically correlated representations and thereafter increase the interpretability of the diagnosis network. Extensive experimental analysis on a private dataset of 466 cases has shown that the proposed method achieves 84.58% on AUC and 84.07% on recall, significantly exceeding the baseline methods.

Automated summary

This paper proposes a Two-stage Expert-guided Diagnosis (TED) framework for MVI in HCC, which predicts key imaging attributes and leverages attention and soft supervision to improve the interpretability of the network.