<![CDATA[Convolutional Neural Networks (CNN) are recognized for their high accuracy in image classification, but large-scale datasets and significant computer resources are needed to train them from scratch, though. Transfer learning offers a practical solution by leveraging pre-trained models to accelerate training even when data is limited. Although CNNs have been widely applied to skin disease classification, specific evaluations of architectures such as ResNet50V2, ResNet152V2, and MobileNetV2 for monkeypox image classification remain scarce. Therefore, this study aims to comprehensively compare the effectiveness and trade-offs of these architectures in detecting monkeypox through transfer learning. The evaluation focuses on balancing accuracy and computational efficiency across stages, including data collection, preprocessing, model design, training, and testing. The dataset, obtained from Kaggle, consists of 2,310 images across four classes: monkeypox, chickenpox, measles, and normal. Transfer learning was implemented using fine-tuned weights from ImageNet. According to the results, ResNet152V2 needed the most training time but had the lowest loss and the greatest validation accuracy (98.28%). ResNet50V2 maintained a good compromise between accuracy (97.84%) and training efficiency, while MobileNetV2 yielded the best overall classification metrics (97.86% for accuracy, precision, recall, and F1-score), indicating strong generalization. These findings highlight the distinct strengths of each model, offering insights into architecture selection based on specific operational constraints and goals.]]>
