JuryFusionNet: a Condorcet's jury theorem-based CNN ensemble for enhanced monkeypox detection from skin lesion images

Accurate and timely diagnosis of monkeypox (Mpox) is vital for outbreak control and effective treatment. However, its skin lesions often resemble those of chickenpox and measles, making visual diagnosis challenging, even for experts. This issue is worsened by the limited availability of large, labeled multiclass datasets, reducing the effectiveness of deep learning models in real-world scenarios. Moreover, single CNN models often suffer from overfitting and bias, leading to suboptimal performance. While combining multiple models can enhance results, simply increasing the voter pool does not ensure higher accuracy. Including multiple models in the voter pool does not always improve accuracy, as traditional ensembles like stacking and majority voting depend on meta-learners or equal weighting, leading to instability and higher computational cost when classifier performance varies. To address these challenges, our study introduces JuryFusionNet, an innovative approach leveraging Condorcet's jury theorem (CDJT) to fuse CNN models. This enhances accuracy while reducing computational complexity by eliminating the need for a meta-learner. First, we leverage transfer learning (TL) and employ four base models, namely DenseNet169, DenseNet201, MobileNet, and ResNet50V2, to extract deep features from skin lesion images. To enhance the performance of these models, we integrate squeeze and excitation (SE) block, which dynamically recalibrate the feature maps to emphasize informative channels and improve model convergence. Furthermore, we explore the power of ensemble learning by applying CDJT, which leverages the collective decision-making of multiple models to improve prediction accuracy through majority voting. Leveraging CDJT, we established that the inclusion of a model in the voter pool enhances the probability of an accurate majority vote, particularly when the model surpasses the accuracy of the other models. We validate our model's effectiveness using two Mpox datasets. The MSID dataset consists of 770 images divided into four classes: chickenpox, measles, Mpox, and normal skin, while the MSLD dataset includes 3192 samples classified as Mpox or non-Mpox. We use Grad-CAM analysis to visualize each base model's attention, confirming the SE block's role in accurately highlighting infected regions. The experimental results demonstrate remarkable accuracy rates of 91.55% on the MSID dataset and 100% on the MSLD dataset, outperforming various pre-trained models, traditional ensemble methods, and existing approaches. Specifically, JuryFusionNet surpasses standard CNN models by 1.94-5.19% and improves upon prior ensemble methods by 0.65-2.59%, underscoring its superior effectiveness in Mpox detection.