Transfer Learning for Medical Waste Image Classification Using EfficientNet-B0 with 5-Fold Cross-Validation

Medical waste management requires careful handling due to its potential to cause infection and environmental hazards when not properly treated. Manual classification of medical waste is time-consuming, highly dependent on human accuracy, and prone to error. This study employs a transfer learning approach using the EfficientNet-B0 architecture to automatically classify medical waste images. The dataset consists of 23 waste categories and undergoes preprocessing steps including data cleaning, resizing, normalization, and data augmentation. The model is initialized with ImageNet-pretrained weights and refined through fine-tuning. Experimental results demonstrate stable classification performance, achieving an average accuracy of 92.2%, precision of 94.1%, recall of 92.2%, and an F1-score of 92.1%. The results indicate that EfficientNet-B0 provides a competitive balance between classification performance and computational efficiency for medical waste image classification, particularly under limited data conditions. However, this study is limited by the size and scope of the dataset and the absence of real-world deployment evaluation, which may affect the generalizability of the results. Despite these limitations, the proposed approach offers a feasible basis for further research toward automated medical waste sorting systems.

Keywords: digital image, EfficientNet-B0, image classification, medical waste, transfer learning