Ensemble of Fine-tuned Deep Learning Models for Monkeypox Detection: A Comparative Study

Abstract

Monkeypox is a rare viral disease that is caused by the monkeypox virus. Monkeypox has clinical symptoms that are similar to those of other diseases such as measles and chickenpox, which makes early detection challenging. The early detection of monkeypox is essential to prevent its spread and reduce the risk of human-to-human transmission. Our study introduces a new method for detecting monkeypox at an early stage using modified transfer learning (TL) algorithms and an ensemble algorithm. The proposed approach can effectively distinguish it from other diseases that have similar symptoms. We used two different datasets, the “Monkeypox Skin Images Dataset (MSID)” and the "Monkeypox-dataset-2022(MD-2022)", which contain images from four classes, including monkeypox, measles, chickenpox, and normal images. We used stratified cross-validation to ensure that each fold of the cross-validation procedure contains a representative sample of each class, which is important when dealing with imbalanced datasets. To evaluate our proposed approach, we trained five pre-trained models, namely DenseNet121, ResNet152V2, ResNet50, InceptionV3, and EfficientNetV2B3, on each dataset separately. The achieved accuracy scores for the MD-2022 dataset were 89.4%, 84.2%, 89.4%, 84.2%, and 84.2%, respectively, while for the MSID dataset, the accuracy scores were 97.4%, 96.2%, 93.6%, 93.6%, and 95% for DenseNet121, ResNet50, InceptionV3, EfficientNetV2B3, and ResNet152V2, respectively. Subsequently, we constructed an ensemble model using a majority voting approach, which combined the predictions of the five models. Our findings indicate that the ensemble model outperformed certain individual models and demonstrated higher efficacy in monkeypox detection by achieving an accuracy score of 89.4% and 98.7% for the "Monkeypox-dataset-2022" and “Monkeypox Skin Images Dataset (MSID)” respectively.