<![CDATA[Eczema (atopic dermatitis) is a chronic inflammatory skin condition whose severity can be difficult to assess consistently from images. Although convolutional neural networks have been applied to dermatological classification, evidence on three-class eczema severity classification using a lightweight transfer-learning model and its implementation as a research prototype remains limited. This study developed an EfficientNetB0 classifier for Normal, Mild, and Severe classes using 1,950 source images. The images were processed at 224 × 224 pixels, augmented only in the training set, and trained in two stages: feature extraction and fine-tuning. The generator produced 1,560 training images and 389 development-validation images. The best model was selected by validation AUC and achieved 89.97% validation accuracy and an AUC of 0.9716. Normal had the highest recall (0.99), while Severe had the highest precision (0.95); most errors involved the Mild and Severe classes. The model was integrated into a browser-based prototype using Flask, HTML, Tailwind CSS, and JavaScript. Because the same validation subset was used for early stopping, checkpoint selection, and final reporting, these results should not be interpreted as independent test performance. The prototype is intended for research and educational use and does not replace professional diagnosis.]]>
Copyrights © 2026
