The improper disposal and mismanagement of medical waste pose severe environmental and public health risks, contributing to greenhouse gas emissions and the spread of infectious diseases. Efficient and accurate medical waste classification is crucial for mitigating these risks. We explore the integration of capsule networks with a pretrained DenseNet model to improve medical waste classification. To the best of our knowledge, capsule networks have not yet been applied to this task, making this study the first to assess their effectiveness. A diverse dataset of medical waste images collected from multiple public sources, is used to evaluate three model configurations: (1) a pretrained DenseNet model as a baseline, (2) a pretrained DenseNet with frozen layers combined with a capsule network, and (3) a pretrained DenseNet with unfrozen layers combined with a capsule network. Experimental results demonstrate that incorporating capsule networks improves classification performance, with F1 scores increasing from 0.89 (baseline) to 0.92 (hybrid model with unfrozen layers). This highlights the potential of capsule networks to address the spatial limitations of traditional convolutional models and improve classification robustness. While the capsule-enhanced model demonstrated improved classification performance, direct comparisons with prior studies were challenging due to differences in dataset size and diversity. Previous studies relied on smaller, domain-specific datasets, which inherently yielded higher accuracy. In contrast, our study employs a significantly larger and more diverse dataset, leading to better generalization but introducing additional classification challenges. This highlights the trade-off between dataset complexity and model performance.

翻译：医疗废物的不当处置和管理会带来严重的环境和公共卫生风险，加剧温室气体排放并助长传染病的传播。高效准确的医疗废物分类对于缓解这些风险至关重要。本研究探索将胶囊网络与预训练的DenseNet模型相结合以改进医疗废物分类。据我们所知，胶囊网络尚未应用于此任务，因此本研究首次评估其有效性。我们使用从多个公共来源收集的多样化医疗废物图像数据集，评估了三种模型配置：(1) 预训练的DenseNet模型作为基线，(2) 冻结层预训练DenseNet与胶囊网络结合，(3) 非冻结层预训练DenseNet与胶囊网络结合。实验结果表明，引入胶囊网络可提升分类性能，F1分数从0.89（基线）提升至0.92（非冻结层混合模型）。这凸显了胶囊网络在解决传统卷积模型空间局限性、提升分类鲁棒性方面的潜力。尽管胶囊增强模型表现出更好的分类性能，但由于数据集规模和多样性差异，与先前研究进行直接比较存在挑战。先前研究依赖规模较小、领域特定的数据集，这本质上会获得更高准确率。相比之下，本研究采用规模显著更大、更多样化的数据集，虽能实现更好的泛化能力，但也带来了额外的分类挑战。这揭示了数据集复杂性与模型性能之间的权衡关系。