Data-Centric Deep Learning for Monkeypox Detection: Hyperdatasets Toward Robust and Generalizable Medical AI

In this work, a structured hierarchy of hyperdatasets for Monkeypox image classification is presented, aiming to evaluate the robustness and generalization ability of deep classification models for Mpox detection. A systematic review is first conducted to identify 11 publicly available Mpox image datasets, while five hyperdatasets are constructed from them. Each hyperdataset represents a distinct level of complexity, spanning predefined and standardized class train-test splits, fully merged heterogenous datasets, and clinically-oriented binary classification datasets. Five state-of-the-art deep architectures are trained on both the original datasets and the five hyperdatasets, aiming to systematically examine how dataset composition, class granularity and domain variability may influence models’ performance, as well as the potential of hyperdatasets towards building more robust and generalizable models. Our results indicate that large-scale heterogenous integration of datasets can provide multiple benefits in the case of limited medical data to further strengthen the diagnostic accuracy of models, especially in clinically valuable binary classification, leading to an accuracy of up to 99.17% with DenseNet-169. In the case of multi-class classification, standardized re-split of train-test data is considered essential after datasets merging, leading to improved classification accuracy and avoidance of classical cross-domain overfitting. Results highlight the importance of hyperdatasets as a methodological approach for developing reliable and generalizable artificial intelligence (AI) systems for supporting infectious skin disease diagnosis.