Kumar Saha Dip, Rafi Sadman, Mridha M F, Alfarhood Sultan, Safran Mejdl, Kabir Md Mohsin, Dey Nilanjan
Department of CSE, Stamford University Bangladesh, Siddeswari, Dhaka, Bangladesh.
Department of CSE, American International University-Bangladesh, Kuratoli, Dhaka, Bangladesh.
BMC Infect Dis. 2025 Mar 25;25(1):403. doi: 10.1186/s12879-025-10811-y.
The daily surge in cases in many nations has made the growing number of human monkeypox (Mpox) cases an important global concern. Therefore, it is imperative to identify Mpox early to prevent its spread. The majority of studies on Mpox identification have utilized deep learning (DL) models. However, research on developing a reliable method for accurately detecting Mpox in its early stages is still lacking. This study proposes an ensemble model composed of three improved DL models to more accurately classify Mpox in its early phases. We used the widely recognized Mpox Skin Images Dataset (MSID), which includes 770 images. The enhanced Swin Transformer (SwinViT), the proposed ensemble model Mpox-XDE, and three modified DL models-Xception, DenseNet201, and EfficientNetB7-were used. To generate the ensemble model, the three DL models were combined via a Softmax layer, a dense layer, a flattened layer, and a 65% dropout. Four neurons in the final layer classify the dataset into four categories: chickenpox, measles, normal, and Mpox. Lastly, a global average pooling layer is implemented to classify the actual class. The Mpox-XDE model performed exceptionally well, achieving testing accuracy, precision, recall, and F1-score of 98.70%, 98.90%, 98.80%, and 98.80%, respectively. Finally, the popular explainable artificial intelligence (XAI) technique, Gradient-weighted Class Activation Mapping (Grad-CAM), was applied to the convolutional layer of the Mpox-XDE model to generate overlaid areas that effectively highlight each illness class in the dataset. This proposed methodology will aid professionals in diagnosing Mpox early in a patient's condition.
许多国家每日新增病例数的激增，使得人类猴痘（Mpox）病例数量不断增加，成为一个重要的全球关注点。因此，尽早识别猴痘以防止其传播至关重要。大多数关于猴痘识别的研究都采用了深度学习（DL）模型。然而，仍缺乏开发一种可靠方法来在早期阶段准确检测猴痘的研究。本研究提出了一种由三个改进的DL模型组成的集成模型，以更准确地在早期阶段对猴痘进行分类。我们使用了广泛认可的猴痘皮肤图像数据集（MSID），其中包括770张图像。使用了增强的Swin Transformer（SwinViT）、提出的集成模型Mpox-XDE以及三个改进的DL模型——Xception、DenseNet201和EfficientNetB7。为了生成集成模型，通过Softmax层、密集层、展平层和65%的随机失活将这三个DL模型进行组合。最后一层中的四个神经元将数据集分为四类：水痘、麻疹、正常和猴痘。最后，实施全局平均池化层以对实际类别进行分类。Mpox-XDE模型表现出色，测试准确率、精确率、召回率和F1分数分别达到98.70%、98.90%、98.80%和98.80%。最后，将流行的可解释人工智能（XAI）技术——梯度加权类激活映射（Grad-CAM）应用于Mpox-XDE模型的卷积层，以生成有效突出数据集中每个疾病类别的叠加区域。所提出的方法将有助于专业人员在患者病情早期诊断猴痘。
