PoxNetX: Viral Pox skin lesion dataset
Description
This dataset provides a curated, multi-source collection of skin lesion images spanning five clinically significant classes: Monkeypox, Chickenpox, Cowpox, Measles, and Normal (healthy) skin. These conditions share overlapping dermatological features, particularly at early lesion stages, making automated differential diagnosis both medically critical and technically challenging. The dataset was assembled as training data for hybrid convolutional neural network (CNN) architectures developed under the PoxNetX research framework at the Department of Information and Communication Technology, Comilla University, Bangladesh. Raw images were aggregated from 14 publicly available source dataset published between 2022 and 2025, hosted across Kaggle, Mendeley Data, and Roboflow repositories. Each source was selected on the basis of image quality, label accuracy, and license compatibility. After aggregation and quality filtering, the curated raw collection comprises 1,875 images distributed as follows: Monkeypox (556), Chickenpox (506), Measles (150), Cowpox (117), and Normal skin (547). Cowpox and Measles are substantially underrepresented, producing a worst-case class imbalance ratio of approximately 4.8 to 1 relative to the dominant classes. Reference keys, publication years, and license strings for all 16 source datasets are documented in the accompanying metadata file. Each source is listed individually in the Related Links section of this record and must be cited independently. To resolve the imbalance, synthetic images were generated for underrepresented classes using the DermGAN-DDPM hybrid generative pipeline, yielding a final balanced dataset with uniform class distribution across all five categories. The released files separate raw curated images from DermGAN-DDPM synthetic outputs, allowing researchers to use either collection depending on their experimental design. Users must consult the source licenses in the metadata before any commercial application, as one upstream source (Ali et al., 2024) carries a CC BY-NC 4.0 non-commercial restriction.
Files
Steps to reproduce
The synthetic images in this dataset were produced using DermGAN-DDPM, a four-stage hybrid generative pipeline designed for skin lesion image synthesis from limited training samples. The pipeline was implemented in Python 3.13 with PyTorch 2.7.1 and trained on an NVIDIA GeForce RTX 3050 Laptop GPU (6 GB GDDR6, CUDA 8.6) using a fixed random seed of 42 for reproducibility. Each stage was applied independently per class, using only that class's available raw images as input. Stage 1: Convolutional Autoencoder. Raw lesion images, resized to 224x224 pixels, were compressed into a structured latent representation by a convolutional autoencoder. The encoder applies successive convolutional blocks with batch normalisation and ReLU activation to reduce spatial dimensionality. The decoder reconstructs images through transposed convolutions. A separate autoencoder was trained per class to ensure each class occupies a distinct region of the shared latent space. Stage 2: Latent Denoising Diffusion Probabilistic Model (DDPM). A DDPM was trained to operate within the compressed latent space from Stage 1 rather than directly in pixel space, substantially reducing computational overhead. The forward diffusion process adds Gaussian noise to latent representations over T timesteps. The reverse process, implemented as a U-Net with residual blocks and self-attention layers, learns to denoise latent samples iteratively. Novel latent samples drawn from the learned distribution were decoded through the Stage 1 autoencoder decoder to produce raw synthetic image candidates. Stage 3: Pix2Pix Conditional Adversarial Refiner. Raw synthetic candidates from Stage 2 were refined using a Pix2Pix image-to-image translation network to enhance texture fidelity, boundary definition, and lesion morphological detail. The generator adopts a U-Net architecture, and the discriminator uses a PatchGAN design that assesses realism at the local patch level rather than globally. Training combined an adversarial loss with an L1 reconstruction loss to balance perceptual quality against structural fidelity to real lesion images. Stage 4: Preprocessing and ROI Annotation. All generated images were normalised using ImageNet channel statistics (mean: [0.485, 0.456, 0.406]; std: [0.229, 0.224, 0.225]). Region-of-interest polygon annotations were applied via the VGG Image Annotator (VIA) to isolate diagnostically relevant lesion regions from background. Only synthetic outputs passing visual quality inspection were retained in the final release. The complete generation code, trained model weights, and annotated notebooks are available in the data/data_generation/ directory of the PoxNetX GitHub repository (see Related Links).
Institutions
- Comilla UniversityChittagong, Comilla