Monkeypox Skin Lesion Dataset

• The Monkeypox Skin Lesion Dataset is a curated binary corpus of lesion images that distinguishes monkeypox from visually similar conditions like chickenpox and measles.
• It employs an extensive manual curation process with quality filtering, cross-referencing, and augmentation to improve image reliability and support transfer learning.
• Benchmark studies using pretrained CNNs and lightweight models underscore its role as a foundational tool for rapid screening and feasibility research.

The Monkeypox Skin Lesion Dataset (MSLD) is an openly accessible skin-lesion image dataset introduced in 2022 to support computer-aided screening of monkeypox from lesion photographs when confirmatory Polymerase Chain Reaction (PCR) testing is not readily available (Ali et al., 2022). Although the source imagery includes monkeypox, chickenpox, and measles cases, the original task formulation is explicitly binary: Monkeypox versus Others, with the latter combining chickenpox and measles because of their visual similarity to monkeypox lesions (Ali et al., 2022). MSLD became one of the earliest public benchmarks for transfer-learning-based mpox lesion recognition and was subsequently reused in studies on lightweight CNNs, explainable AI, edge deployment, and few-shot learning (Dihan et al., 2024, Delgado-López et al., 23 Jul 2025, Rashid et al., 6 May 2026).

1. Origin and intended scope

MSLD was created because, at the time of the original study, there was no publicly available monkeypox skin-lesion dataset suitable for developing automated detection methods (Ali et al., 2022). The motivating clinical problem was early differential diagnosis: monkeypox lesions can resemble chickenpox and measles, and this similarity complicates rapid recognition, especially in settings where PCR confirmation is inaccessible or delayed (Ali et al., 2022).

The dataset was intended for binary classification, screening support, and transfer-learning research on a small medical image corpus (Ali et al., 2022). In the original feasibility study, the authors positioned MSLD as a basis for early identification of suspected monkeypox cases and for eventual deployment in a web-based screening tool (Ali et al., 2022).

This framing is important because MSLD was never presented as a definitive diagnostic resource. Its role was preliminary and infrastructural: to establish whether transfer learning on lesion photographs could support surveillance and rapid triage during an outbreak. A plausible implication is that MSLD should be interpreted primarily as an early benchmark for feasibility studies rather than as a mature clinical dataset.

2. Acquisition, verification, and curation pipeline

MSLD was manually compiled from publicly available case reports, news portals, and websites, and the paper explicitly states that the authors did not use automatic web-scrapers (Ali et al., 2022). Instead, the collection process relied on extensive manual searching in order to preserve image quality and label reliability (Ali et al., 2022). The retained images came from patients with lesions on body regions including the face, neck, hand, arm, and leg (Ali et al., 2022).

To improve reliability, all images were checked using Google Reverse Image Search and were cross-referenced with other sources (Ali et al., 2022). Quality control then proceeded through a 2-stage screening process that removed out-of-focus images, low-resolution images, and low-quality images; only unique images that satisfied the quality criteria were retained (Ali et al., 2022). The selected images were then cropped to the region of interest, resized to 224 × 224 pixels, and processed while maintaining aspect ratio (Ali et al., 2022). The dataset was made openly accessible through the project GitHub repository mentioned in the paper (Ali et al., 2022).

These design choices distinguish MSLD from a raw web scrape. The dataset is still web-sourced, but its curation pipeline includes source triangulation, duplicate control, quality filtering, and lesion-focused preprocessing. That combination made it unusually usable for CNN experimentation relative to more heterogeneous outbreak-era image collections.

3. Class structure, scale, and augmentation regime

MSLD contains two labels: Monkeypox and Others, where Others combines chickenpox and measles (Ali et al., 2022). The original dataset contains 228 images from 162 unique patients: 102 monkeypox images from 55 patients and 126 others images from 107 patients (Ali et al., 2022).

Class label	Original images	Augmented images
Monkeypox	102	1428
Others	126	1764
Total	228	3192

Because monkeypox images were scarce, the dataset was expanded by about 14-fold through augmentation (Ali et al., 2022). The augmentation methods listed in the paper are rotation, translation, reflection, shear, hue jitter, saturation jitter, contrast jitter, brightness jitter, noise, and scaling (Ali et al., 2022). A figure in the paper further illustrates 13 augmented versions of one original monkeypox image, including random rotation by a multiple of 90°, random rotation in the range of −45° to 45°, translation, reflection, shear, hue jitter, saturation jitter, brightness jitter, contrast jitter, salt-and-pepper noise, Gaussian noise, synthetic blur, and scaling (Ali et al., 2022).

The resulting augmented corpus contains 1,428 monkeypox images and 1,764 others images, for a total of 3,192 images (Ali et al., 2022). The original dataset is therefore slightly imbalanced, both before and after augmentation, in favor of the Others class (Ali et al., 2022).

A common misunderstanding is that MSLD is intrinsically a large dataset. It is not. The large image count is augmentation-derived; the original evidence base remains 228 images from 162 patients. This distinction is central to interpreting both model performance and generalization claims.

4. Original experimental protocol and baseline benchmarking

The original feasibility study evaluated MSLD with 3-fold cross-validation and a split into training, validation, and test sets with an approximate ratio of 70:10:20, while maintaining patient independence (Ali et al., 2022). An important design decision was that only the training and validation sets were augmented, whereas the test set contained only original images (Ali et al., 2022). This prevented test-time evaluation on synthetic samples.

The study benchmarked three ImageNet-pretrained CNNs—VGG-16, ResNet50, and InceptionV3—together with an ensemble model based on majority voting (Ali et al., 2022). The input size was 224 × 224 × 3; the pretrained fully connected layers were removed; the bottom 8 layers were unfrozen; and a new classifier head was added with FC 4096, FC 1072, FC 256, and a final FC layer with 2 outputs + softmax (Ali et al., 2022). The reported dropout rates were 0.3, 0.2, and 0.15; the optimizer was Adam with learning rate 1e−5; batch size was 16; and the loss was binary cross-entropy (Ali et al., 2022).

The reported 3-fold results were as follows: VGG16 achieved 81.48 ± 6.87 accuracy, 0.85 ± 0.08 precision, 0.81 ± 0.05 recall, and 0.83 ± 0.06 F1; ResNet50 achieved 82.96 ± 4.57 accuracy, 0.87 ± 0.07 precision, 0.83 ± 0.02 recall, and 0.84 ± 0.03 F1; InceptionV3 achieved 74.07 ± 3.78 accuracy, 0.74 ± 0.02 precision, 0.81 ± 0.07 recall, and 0.78 ± 0.04 F1; and the ensemble achieved 79.26 ± 1.05 accuracy, 0.84 ± 0.05 precision, 0.79 ± 0.07 recall, and 0.81 ± 0.02 F1 (Ali et al., 2022). ResNet50 was therefore the best overall model in the original benchmark (Ali et al., 2022). The same study also produced a prototype web-application as an online monkeypox screening tool (Ali et al., 2022).

These numbers established MSLD as a feasibility benchmark rather than a saturated one. The gap between ResNet50 and InceptionV3, and the fact that the ensemble did not surpass the best single model, indicate that representation choice mattered strongly under low-data conditions.

5. Subsequent reuse as a benchmark

Later work reused MSLD in several distinct methodological settings. MpoxSLDNet treated the augmented MSLD as a binary benchmark of 3,192 images, used a 90% training / 10% testing/validation split, and reported 94.56% validation/testing accuracy for a lightweight custom CNN, compared with 86.25% for VGG16, 84.38% for DenseNet121, and 67.19% for ResNet50 under that paper’s protocol (Dihan et al., 2024). An Attention Based Feature Fusion Network later used the augmented Kaggle MSLD with 4-fold cross-validation and reported 96.52% mean validation accuracy, 96.58% precision, 96.52% recall, and 96.51% F1-score (Kundu et al., 2024).

MSLD was also adopted in deployment-oriented studies. An embedded-system study trained MobileNetV2 on the open-source MSLD for binary classification on the NVIDIA Jetson Orin Nano and reported 92.19% accuracy, 91.30% precision, 94.92% recall, and 93.07% F1-score (Delgado-López et al., 23 Jul 2025). That same paper noted an important failure mode: on a 320-image test set, 20 images were misclassified and 15 of those errors were monkeypox images mislabeled as Others, highlighting a false-negative risk in screening applications (Delgado-López et al., 23 Jul 2025).

The dataset has additionally served as a protocol stress test. The ITMAINN study used the Kaggle version of the binary MSLD alongside stratified 80:20 train-test split and stratified 5-fold cross-validation, again emphasizing that augmentation was applied only to the training data (Alghoraibi et al., 25 May 2025). A later few-shot benchmark used the original, unaugmented images only from MSLD v1.0, defining it as a 2-way benchmark with 1-shot, 5-shot, and 10-shot episodes (Rashid et al., 6 May 2026).

Taken together, these studies show that MSLD has functioned less as a single fixed benchmark than as a reusable binary corpus for transfer learning, architecture comparison, compression, embedded deployment, explainability, and few-shot representation testing. This suggests that reported accuracies on MSLD should be interpreted in light of the specific split policy, augmentation regime, and evaluation protocol used by each study.

6. Limitations, misconceptions, and relation to later datasets

The original paper identifies several limitations: a small number of unique patients; limited demographic diversity in geography, race, and gender; absence of richer metadata such as clinical history, days since symptom onset, and lesion stage; labels based on web-sourced material rather than prospectively collected clinical data; and pretraining on ImageNet, which does not contain skin-lesion images (Ali et al., 2022). The authors recommend a larger, more diverse dataset, improved demographic coverage, inclusion of clinical metadata, use of a multi-source dermatoscopic pretraining dataset, and expansion through international collaboration (Ali et al., 2022).

A recurrent misconception is to conflate the original MSLD with its later follow-up. The original MSLD is a binary dataset with Monkeypox and Others. The expanded MSLD v2.0 is a distinct six-class dataset containing mpox, chickenpox, measles, cowpox, hand-foot-mouth disease (HFMD), and healthy, with 755 images from 541 distinct patients and dermatologist verification of disease labels (Ali et al., 2023). MSLD v2.0 was created precisely to address limitations in the earlier release, including limited class coverage, image quality issues, and skin-tone representation (Ali et al., 2023).

The broader literature has treated these limitations as characteristic of the field rather than of MSLD alone. A 2023 survey describes the mpox skin-lesion dataset ecosystem as constrained by limited, non-standardized, and often small-scale datasets, with recurring issues of class imbalance, label uncertainty, duplication risk, and limited external validity (Khan et al., 2023). Within that landscape, MSLD remains historically important because it filled an early public-data gap and established a concrete benchmark for lesion-based monkeypox screening. Its lasting significance lies less in its size than in its role as a catalyst for subsequent dataset releases, protocol refinements, and methodological comparisons across the mpox computer-vision literature.