Explainable HCC Diagnosis on Dynamic Contrast-Enhanced MRI with a Li-RADS Concept Bottleneck

Killian Monnin, Patrick Jeltsch, Lucia Fernandes-mendes, Vasco Cazzagon, Murat Yüce, Vivek Yadav, Mario Jreige, Marianna Gulizia, Montserrat Fraga Christinet, Raphaël Girardet, Clarisse Dromain, Bachir Taouli, Naïk Vietti-Violi, Jonas Richiardi

Keywords: HCC, Explainable AI, Classification, MRI, Concept bottleneck

TL;DR: Li-RADS Concept Bottleneck for Explainable HCC MRI

Abstract: We propose an explainable end-to-end framework for hepatocellular carcinoma (HCC) diagnosis on dynamic contrast-enhanced (DCE) liver MRI. Our method embeds Liver Imaging Reporting and Data System (Li-RADS)–inspired concepts into the network via a multi-head concept bottleneck. A 2.5D EfficientNet backbone processes lesion-centred multiphase MRI crops, and a 4-head architecture jointly predicts continuous soft labels for non-rim arterial phase hyperenhancement (APHE), portal venous/delayed washout and capsule, lesion morphology, and a LR-5 score (definite HCC vs non-HCC) based on the Li-RADS guidelines. Soft labels are derived automatically from intra-lesional, peri-lesional and parenchymal intensity patterns, and the network is trained with uncertainty-weighted losses to balance concept prediction, contrast regression and HCC classification. On our cohort, the 4-head Li-RADS–inspired bottleneck substantially improves Grad-CAM explanation accuracy, geometric stability and intensity robustness while maintaining PR AUC comparable to a single-head baseline, highlighting an interpretable alternative to a black-box HCC classifier.

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 304