AICOM-MP: an AI-based monkeypox detector for resource-constrained environments

Tim Tianyi Yang, Tom Tianze Yang, Andrew Liu, Na An, Shaoshan Liu, Xue LiuConnect. Sci 2022

Introduction

Monkeypox is a highly contagious skin disease that has become a significant global health concern. In Least Developed Countries (LDCs), clinical diagnosis via Polymerase Chain Reaction (PCR) is often prohibitively expensive or physically inaccessible. While AI-based detectors offer a scalable alternative, existing models frequently fail in real-world scenarios. They are typically trained on high-quality lab datasets that do not reflect the "noisy" backgrounds or low-resolution sensors characteristic of budget mobile phones used in LDCs.

To bridge this "digital divide," the authors present AICOM-MP, a state-of-the-art (SOTA) monkeypox detector developed under the Autonomous Mobile Clinics (AMCs) initiative. AICOM-MP is designed to deliver high-accuracy diagnostics using low-end mobile devices in complex, resource-constrained environments.

The Autonomous Mobile Clinic (AMC) ecosystem integrating AICOM-MP for remote diagnostics.

Methods: A Hierarchical Vision Pipeline

The core innovation of AICOM-MP is a three-level vision pipeline that mimics a physician's visual diagnostic process. By progressively narrowing the model's focus, the system reduces false positives caused by irrelevant background objects.

  1. Level 1: Human-Object Segmentation ($U^2$-Net): This initial layer identifies the human subject and removes background noise such as furniture or outdoor scenery.
  2. Level 2: Region-based Skin Detection (FCNResNet10): The model then isolates skin regions, pruning away non-skin features like clothing or jewelry.
  3. Level 3: Deep-Learning Classification (EfficientNetB7): Finally, the model analyzes isolated skin patches for monkeypox lesions. EfficientNetB7 was selected for its optimal performance-to-compute ratio, making it suitable for mobile CPUs.
The hierarchical AICOM-MP architecture featuring human-object segmentation, skin segmentation, and deep-learning classification.

Additionally, a Resolution Restoration unit pre-processes low-resolution images to ensure that hardware limitations do not compromise diagnostic accuracy.

Datasets and Evaluation

Recognizing that existing datasets were often small or lacked diversity, the authors introduced two new benchmarks:

  • AICOM-MP Training Dataset: A balanced set of 4,932 images covering monkeypox, healthy skin, and other diseases.
  • COCO_MP Dataset: A challenging test set of 332 images designed to evaluate generalization in complex, real-world environments.

Key Results

AICOM-MP achieved 100% accuracy on binary classification tasks using the precedent Ali et al. dataset. More importantly, on the challenging COCO_MP dataset, the full hierarchical pipeline maintained a high accuracy of 96.34%. Ablation studies confirmed that the background removal and skin segmentation steps were critical, improving accuracy from 88.55% to over 96%.

Conclusion and Future Work

AICOM-MP demonstrates that "context cleaning" is as vital as the classifier itself for real-world medical AI. While currently a web service, future work focuses on migrating the model to run fully offline on low-end devices and expanding its diagnostic breadth to distinguish between similar conditions like chickenpox and measles.

Summary updated

BibTeX
@inproceedings{yang2022aicom,
  author = {Tim Tianyi Yang and Tom Tianze Yang and Andrew Liu and Na An and Shaoshan Liu and Xue Liu},
  title = {{AICOM-MP: an AI-based monkeypox detector for resource-constrained environments}},
  booktitle = {Connect. Sci},
  year = {2022},
}
Figure