MicroMPoxNet

MicroMPoxNet: A Lightweight CNN for Interpretable Monkeypox Classification in Dermoscopic Images with Comparative Transfer Learning Evaluation

Authors
A. A. M. Mustahid, A. A. M. Muzahid*, Md. Sadekur Rahman, Hua Han, Yujin Zhang, Ferdous Sohel

📄 Status: Published

---

Repository Structure

MicroMPoxNet/
│
├── README.md
├── LICENSE
├── CITATION.cff
│
├── assets/
│   ├── pipeline.png
│   ├── architecture.png
│   ├── dataset_samples.png
│   ├── TLModel_Diagram.jpg
│   ├── XAI_n.jpg
│   ├── conf_roc_curve.png
│   ├── model_performance_tradeoffs.png
│   └── Final_JournalGraph.png
│
├── models/
│   └── description.md
│
└── dataset/
    └── dataset_links.md

---

Table of Contents

• Overview
• Highlights
• Research Pipeline
• Architecture Overview
• Dataset
• Explainability (XAI)
• Code Availability
• Citation
• Contact

---

Research Pipeline

---

Overview

Monkeypox (Mpox) has recently emerged as a global public health concern, highlighting the need for fast, accessible, and reliable diagnostic tools. Traditional diagnostic methods such as PCR testing require specialized laboratory infrastructure and trained personnel, which can limit timely diagnosis in resource-constrained environments.

Recent deep learning approaches have shown promising results for automated skin lesion classification. However, many existing methods rely on large and computationally expensive transfer learning models, making real-time deployment difficult in practical healthcare settings.

To address these limitations, this work introduces MicroMPoxNet, a lightweight convolutional neural network designed for efficient and interpretable Mpox classification from dermoscopic images. The proposed architecture integrates depthwise separable convolutions, residual connections, Swish activation functions, and squeeze-and-excitation blocks to achieve a strong balance between predictive performance and computational efficiency.

Extensive experiments were conducted using both transfer learning baselines and the proposed MicroMPoxNet model. The results demonstrate that MicroMPoxNet achieves highly competitive performance while maintaining significantly lower computational complexity, making it suitable for deployment in low-resource clinical environments.

---

Highlights

• Proposed MicroMPoxNet, a lightweight CNN architecture for Monkeypox skin lesion classification.
• Designed for efficient deployment in resource-constrained clinical environments.
• Integrates depthwise separable convolutions, residual connections, Swish activation, and squeeze-and-excitation blocks.
• Achieves high classification performance with significantly reduced computational complexity.
• Evaluated against 11 state-of-the-art transfer learning models.
• Includes explainable AI methods (Grad-CAM, LIME, Integrated Gradients) for improved clinical interpretability.

---

Architecture Overview

MicroMPoxNet is designed with the following goals:

• Efficient Mpox skin lesion classification from dermoscopic images
• Lightweight architecture suitable for real-time deployment in resource-constrained environments
• Reduced computational complexity while maintaining strong predictive performance
• Improved interpretability for clinical decision support

The architecture emphasizes efficient feature extraction and compact network design to ensure reliable Mpox classification while enabling deployment on low-resource medical devices and edge computing platforms.

---

Architecture of MircoMPoxNet

Architecture of Transfer Learning

Classification results on Binary and Multi-class dataset

Model Performance Tradeoffs (MicroPoxNet vs Transfer Learning)

---

Dataset

Mpox2025 Dataset (Proposed)

The Mpox2025 dataset was constructed to support reliable Mpox classification using dermoscopic images.

Dataset characteristics:

• Total images: 696
• Mpox images: 337
• Normal images: 359
• Image format: PNG
• Color space: RGB / Grayscale

After applying data augmentation techniques, the dataset was expanded to 34,555 images for robust model training.

Sample Images from Mpox2025

Dataset Download

The Mpox2025 dataset used in this study is available at:

Google Drive:
https://drive.google.com/file/d/1P3ZZ01TSJS4v9uhBy0USzfbMXWbO8QIh/view?usp=sharing

External dataset used in this study:

Kaggle Dataset:
https://www.kaggle.com/datasets/maxmelichov/monkeypox-2022-remastered

---

Explainability (XAI)

To improve clinical interpretability, this study employs several explainability techniques:

• Grad-CAM
• LIME
• Integrated Gradients

These methods help visualize which regions of the dermoscopic images contribute most to the model’s predictions, supporting more transparent AI-assisted diagnosis.

---

Code Availability

The complete implementation of MicroMPoxNet, including model architecture and training scripts, will be made publicly available upon acceptance of the paper.

The Mpox2025 dataset used in this study is currently accessible via the Google Drive link provided in the dataset section.

---

Citation

If you find this work useful for your research, please consider citing:

@article{Mustahid2026MicroMPoxNet,
  title = {MicroMPoxNet: A lightweight CNN for interpretable Monkeypox classification in dermoscopic images with comparative transfer learning evaluation},
  author = {Mustahid, A. A. M. and Muzahid, A. A. M. and Rahman, Md. Sadekur and Han, Hua and Zhang, Yujin and Sohel, Ferdous},
  journal = {Biomedical Signal Processing and Control},
  volume = {123},
  pages = {110404},
  year = {2026},
  month = sep,
  issn = {1746-8094},
  doi = {10.1016/j.bspc.2026.110404},
  url = {https://www.sciencedirect.com/science/article/pii/S1746809426009584},
  keywords = {Mpox classification, Dermoscopic images, Explainable AI, Medical image analysis, Skin disease detection},
  note = {Open access}
}

---

Contact

For questions or collaboration inquiries, please contact:

A. A. M. Mustahid
Email: mustahid34@gmail.com

---