DR-CapsNet: deep residual capsule network with dynamic routing for automated identification of hepatocellular carcinoma and cirrhosis in CT images

Abstract

Hepatocellular carcinoma (HCC) diagnosis in CT images is challenging due to the overlapping imaging features with cirrhosis, particularly in early-stage small nodules, where rapid differentiation requires both morphological sensitivity and computational efficiency. To address this, we propose DR-CapsNet, a deep residual capsule network that combines lightweight residual blocks with dynamic routing mechanisms. The residual module alleviates gradient degradation through skip connections while enhancing the extraction of high-level image features, whereas the capsule framework leverages vector neurons and dynamic routing to model hierarchical part-whole relationships between cirrhotic nodules and HCC lesions. The dynamic routing mechanism iteratively refines coupling coefficients to establish affine-invariant spatial correlations across multi-scale capsule layers, allowing for precise differentiation of subtle morphological variations. Experimental results indicate that DR-CapsNet outperforms existing state-of-the-art methods in both accuracy and inference speed. Moreover, it exhibits exceptional robustness, even under limited training conditions (400 samples) and class imbalance (1:4 ratio) challenges. Overall, DR-CapsNet presents an accurate, efficient, and robust solution for the diagnosis of hepatocellular carcinoma (HCC), particularly in clinical settings with constrained resources.