Guiding the classification of hepatocellular carcinoma on 3D CT-scans using deep and handcrafted radiological features

TL;DR

The study tackles non-invasive histology-proven HCC diagnosis from CT by addressing radiologist variability and small lesion challenges. It introduces a LI-RADS-guided two-step framework that first predicts the three LI-RADS major features with 3D CNNs and then fuses these predictions with lesion size and handcrafted radiomics features via logistic regression to predict HCC. The approach achieves significant performance gains over pure deep-learning baselines (AUC improvements of $6$ to $18$ points), with the best fusion reaching an AUC of approximately 0.75 on a challenging dataset and showing favorable results on a private test set, outperforming non-expert radiologists and matching expert performance in some cases. These results highlight the value of integrating radiologists’ criteria through handcrafted features with DL-based feature learning to improve robustness in small-lesion HCC characterization and pave the way for more interpretable, clinically actionable diagnostic tools.

Abstract

Hepatocellular carcinoma is the most spread primary liver cancer across the world ($\sim$80\% of the liver tumors). The gold standard for HCC diagnosis is liver biopsy. However, in the clinical routine, expert radiologists provide a visual diagnosis by interpreting hepatic CT-scans according to a standardized protocol, the LI-RADS, which uses five radiological criteria with an associated decision tree. In this paper, we propose an automatic approach to predict histology-proven HCC from CT images in order to reduce radiologists' inter-variability. We first show that standard deep learning methods fail to accurately predict HCC from CT-scans on a challenging database, and propose a two-step approach inspired by the LI-RADS system to improve the performance. We achieve improvements from 6 to 18 points of AUC with respect to deep learning baselines trained with different architectures. We also provide clinical validation of our method, achieving results that outperform non-expert radiologists and are on par with expert ones.

Figures (4)

• Figure 1: Schema of the LI-RADS five main visual features. The threshold growth, which is obtained with longitudinal scans, is not available in the datasets used in this study.
• Figure 2: Examples of center slices extracted from $\mathcal{D}_1$.
• Figure 3: Overview of the proposed method. $s$ corresponds to the lesion diameter (in mm). The architecture leads to a one-dimensional vector of 8 components, being either a handcrafted feature or a predicted probability of presence of each criterion.
• Figure 4: First row: AUC metric for each deep learning model of LI-RADS major features. Second row: absolute values coefficients of the LI-RADS DLF in the logistic regression, for three backbones.