"ScabAtlas: A Comprehensive Medical Image Dataset for Scabies Detection"

"Scabies, a neglected tropical disease caused by Sarcoptes scabiei var. hominis, remains a worldwide public health issue, particularly in low and middle-income countries, where overcrowding and poor hygiene practices increase the risk of rapid transmission. Therefore, early and accurate diagnosis is essential for preventing outbreaks, yet clinical identification is often proven to be difficult for overlapping syndromes with other dermatological conditions or diseases. In this regard, deep learning-based image analysis offers a promising solution for scalable and automated scabies detection. In this study, we constructed a dedicated Scabies Image Dataset, ScabAtlas, sourced from DernBET, medical camps, and hospital records from Bangladesh, and further augmented it. Using this dataset, we evaluated seven ImageNet-pretrained convolutional neural network architectures, including ResNet18, ResNet50, MobileNetV3 (Small and Large), EfficientNetB0, ConvNeXt-Tiny, and VGG16-BN, using a five-fold cross-validation scheme. Experimental results exhibited that ResNet50 achieved the best overall performance with an accuracy of 88.7% (\u00b10.02) and a macro-F1 score of 0.91, outperforming other models. Grad-CAM visualizations confirmed that the network consistently focused on lesion-affected regions, while embedding-based image retrieval highlighted class-consistent feature representations. We have also developed a mobile application based on the best-performing model that can identify scabies skin lesions, aggregating symptoms with the model's learned features from the images."

疥疮（Scabies）是一种由人疥螨（Sarcoptes scabiei var. hominis）引发的被忽视热带病，至今仍是全球性公共卫生挑战，在中低收入国家尤为突出——这类地区普遍存在人群拥挤、卫生条件欠佳的情况，大幅提升了疾病快速传播的风险。因此，早期精准诊断对于遏制疫情暴发至关重要，但由于其临床症状与多种皮肤病症存在重叠，临床鉴别往往颇具难度。有鉴于此，基于深度学习的图像分析技术为实现规模化、自动化的疥疮检测提供了极具前景的解决方案。本研究构建了专用疥疮图像数据集ScabAtlas，数据来源涵盖孟加拉国的DernBET平台、医疗营地及医院病历，并对数据集进行了扩充增强。本研究采用五折交叉验证方案，基于该数据集对7种在ImageNet上预训练的卷积神经网络架构开展性能评估，涉及模型包括ResNet18、ResNet50、MobileNetV3（小型版与大型版）、EfficientNetB0、ConvNeXt-Tiny及VGG16-BN。实验结果表明，ResNet50取得了最优综合性能，准确率达88.7%（±0.02），宏F1值为0.91，性能优于其余所有参评模型。Grad-CAM可视化结果证实，该网络始终聚焦于皮损受累区域；而基于嵌入的图像检索则凸显了类别一致的特征表征。此外，本研究基于性能最佳的模型开发了一款移动应用，可识别疥疮皮肤皮损，并将临床症状与模型从图像中学习到的特征进行整合应用。