Comparative Evaluation of MoE and HMoE for Multiclass Classification in VCE Image Analysis

Video Capsule Endoscopy (VCE) is a pivotal technology in modern gastroenterology, offering a non-invasive method to visualize the entire small bowel. However, the clinical application of VCE is hampered by the extensive review time required, as specialists must manually analyze thousands of images from each procedure. This process is not only laborious and costly but also prone to diagnostic errors due to fatigue, subtle abnormalities, and variability in interpretation across clinicians. To address this challenge, deep learning methods have been explored to automate VCE image analysis. However, most existing approaches rely on a single model architecture, which often fails to generalize across the broad visual diversity found in gastrointestinal imagery. This limitation becomes especially pronounced in multiclass classification tasks, where the ability to distinguish between visually similar tissues and lesions is essential. Ensemble-based methods such as Mixture of Experts (MoE) have shown promising results in general computer vision by leveraging multiple specialized models for improved robustness. However, no prior work has investigated MoE or Hierarchical MoE (HMoE) architectures for multiclass classification of VCE or endoscopic images more broadly. To explore this opportunity, we present a comparative framework evaluating three deep learning strategies for VCE image classification: individual models, flat MoE systems, and Hierarchical MoE architectures. Using a subset of the Kvasir-Capsule dataset, which contains 12 gastrointestinal tissue and lesion classes, we first train and evaluate four backbone models (InceptionNeXt, EfficientViT, ConvNeXtV2, and DeiT3) to establish a performance baseline. The two best-performing architectures, ConvNeXtV2 and DeiT3, are then used as expert backbones within both MoE and HMoE systems. In the MoE configuration, a gating network assigns dynamic per-image weights to multiple expert instances. In contrast, the HMoE configuration constructs a learned binary tree that routes samples based on class similarity through increasingly specialized branches. In the HMoE models, ConvNeXtV2 outperformed DeiT3 in accuracy, whereas DeiT3 showed superior routing accuracy. These results indicate that expert-driven ensemble methods not only outperform standalone models but also offer complementary advantages depending on architecture and routing strategy. This study provides new evidence for the clinical potential of MoE and HMoE frameworks in scalable, accurate VCE image analysis