BIF: A Biosignature Identification Framework for Model-agnostic Interpretation of MVI Diagnosis Models in HCC

Pengyu Zheng, Yong Yi, Ye Luo

Abstract: Microvascular invasion (MVI) is a critical determinant that substantially influences the postoperative prognosis of hepatocellular carcinoma (HCC). Accurate preoperative diagnosis of MVI using MRI imaging has far-reaching research implications. While deep learning models have demonstrated remarkable diagnostic performance, their intrinsic black-box nature poses significant challenges to further advancement. To address this limitation, we propose a novel, model-agnostic interpretation approach, the Biosignature Identification Framework (BIF), inspired by causal inference theory and the biological concept of biosignatures. Within BIF, the Biosignature Identification Module (BIM) operates in parallel with the prediction model, identifying key biosignatures and generating interpretations based on these biosignatures. Unlike conventional model-agnostic interpretation techniques, BIF uniquely offers definitive interpretations grounded in causal inference, thereby enhancing the accuracy and credibility of the interpretive process. Extensive experiments on a clinical dataset collected by Zhongshan Hospital demonstrate the interpretability and efficacy of BIF in preoperative MVI prediction for HCC.