<p>Due to late diagnosis, biological heterogeneity, and limited therapeutic options, hepatocellular carcinoma (HCC) remains one of the most substantial causes of cancer-related mortality. Old diagnosis mechanisms tend to lack enough sensitivity and specificity due to the biological heterogeneity of different liver cancers, especially in the early stages. In this paper, we leverage an approach named as TEXNet-LipTree, which is a hybrid deep learning architecture for robust, interpretable and generalizable hepatocellular carcinoma (HCC) classification on the TCGA-LIHC dataset. The main leverage we employ on this approach is the fact that it combines spatial feature extraction, constrained neural mapping, and symbolic decision making. First, we transform the high dimensional tabular omics data into 2D pseudo-image formats, which allow us to use the Gray-Level Co-occurrence Matrix (GLCM) to extract the second order texture features: contrast, energy, homogeneity, and correlation. These extracted features are then processed through a Lipschitz constrained neural network with Learnable Linear Spline (LLS) activations in order to ensure the model is robust against noise, adversarial perturbations, and interpretability through second-order Total Variation (TV2) mathematical regularization. The final stage is through the decision trees trained on the extracted feature space, which promotes the advantages of being traceable, interpretable, and surgeon-auditable. Experiments performed on the TCGA-LIHC dataset reveal superior performance in both binary and multiclass classification: The binary classification of tumor versus normal tissue 98.89% 98.32%, and the multiclass staging is between 95.75% and 97.77%, with AUC between 95.85% and 97.68%, both outperforming the existing methods such as SVM, CNN, and traditional DNNs.</p>