An Attention-based Deep Learning Approach on Augmented Data for Classifying Monkeypox, Chickenpox, and Measles

Skin disease is a widespread issue that has affected individuals since ancient times. In

today’s modern era, significant advancements have been made in treating skin diseases;

however, these developments are still insufficient to effectively prevent the global spread

of such conditions or to identify them in their early stages. Early detection of skin disease

may help us get better treatment and cure it early. Early skin disease identification is now

easy because of developments in AI and image processing techniques. The affected rates

of chickenpox, measles, and monkeypox have increased significantly over the past few

years. Due to their similar symptoms, medical facilities lack the necessary equipment to

detect these skin diseases in their early stages. For the detection of chickenpox, measles,

and monkeypox, we have used the Monkeypox Skin Images Dataset (MSID), which con-

sists of 770 images. Given the inadequate number of images in the dataset and the pres-

ence of class imbalance issues, we implemented the CycleGAN-based data augmentation

approach. By adopting data augmentation, we have created a dataset of 2000 images, and

each class of the dataset has exactly 500 images. Using the CBAM-DenseNet201 model,

we were able to classify the pox virus with 96.5% test accuracy after removing the class

imbalance issue. Convolutional Block Attention Module (CBAM) was also used as an

attention mechanism to boost the accuracy of the models. We also used other pre-trained

models such as MobileNetv2, Inceptionv3, Xception, and DenseNet169 for experimen-

tal purposes. All of these models, when applied to the augmented data, incorporate the

CBAM attention mechanism. Precision, recall, F1-score, and confusion matrix were uti-

lized to evaluate the model’s performance. As our model works well in detecting skin

disease, it will be beneficial for future use in clinical research.

• Supervision