AICOM-MP: AI Monkeypox Detector

• AICOM-MP is an AI-based monkeypox detector providing standardized, modular health screening for resource-constrained environments.
• It employs a hierarchical, attention-driven visual processing pipeline that combines segmentation and classification to achieve state-of-the-art diagnostic accuracy.
• Designed for both online and offline use, it seamlessly integrates with autonomous mobile clinics through containerized microservices and open-source innovation.

AICOM-MP is an AI-based monkeypox detector developed as part of the Autonomous Mobile Clinics (AMCs) initiative to provide accessible and standardized health AI screening in resource-constrained environments, particularly Least Developed Countries (LDCs). Designed for both online and offline modalities, AICOM-MP offers state-of-the-art performance in monkeypox detection and exemplifies the deployment of domain-specific health AI engines as modular, integrable applications (“apps”) within the AMC platform architecture. It is open-sourced, available as a web service, and includes a diverse, expert-labeled dataset to enable both clinical research and practical deployment in the field (Yang et al., 2022).

1. Position within the Autonomous Mobile Clinic Framework

AICOM-MP is conceptualized as a “health AI engine” that functions analogously to an application on a smartphone, but situated within the AMC hardware-software stack. The AMC platform comprises a self-driving vehicle chassis, medical diagnostic instruments, an edge AI computation node, and a containerized software system. Health AI engines such as AICOM-MP are containerized microservices, each with a dedicated REST API for data ingest and screening tasks. Clinic operators can configure the set of active AI engines according to epidemiological needs by swapping containers, enabling outbreak-specific deployments (e.g., monkeypox, malaria, diabetic retinopathy) (Yang et al., 2022).

Deployment environments range from rural clinics to disaster zones, supporting both continuous and intermittent connectivity through online (web service) and offline (local inference) modes. The backbone software utilizes Linux, Docker/Kubernetes orchestration, ONNX Runtime or TensorFlow-Serving for inference, and integrates with electronic medical records and telemedicine peripherals (Yang et al., 2022).

2. Model Architecture and Inference Pipeline

The AICOM-MP technical pipeline applies a hierarchical, attention-driven visual processing architecture, reflecting the stepwise procedure of human clinical inspection:

• Level-1 (What): Human-object segmentation using U²-Net for background removal while retaining the patient silhouette, avoiding excessive masking (>87% removal).
• Level-2 (Where): Skin region segmentation with FCNResNet10, trained on COCO skin annotations, to filter out irrelevant areas (clothing, rings).
• Level-3 (Which): Lesion classification using an EfficientNet-B7 backbone pre-trained on ImageNet, with a custom head including BatchNorm, a dense layer (L2 kernel regularizer λ=0.016, L1 activity λ=0.016, L1 bias λ=0.006), dropout (rate=0.45), and a two-class softmax.

An optional super-resolution unit (e.g., Deep Latent Space Translation) may be applied to enhance low-resolution input. The preprocessing pipeline accepts JPEG or PNG images from low-end devices, standardizes input to 224×224 pixels, and normalizes pixel values. The final output is a risk label (monkeypox/other) with an accompanying confidence score (Yang et al., 2022).

3. Training Procedure and Dataset Composition

The model is trained on a composite, expert-curated dataset designed for clinical diversity and robustness:

• Training set: 4932 images, balanced post-augmentation between monkeypox and non-monkeypox classes.
• Validation set: 234 images (split equally).
• Test set: 126 images (split equally).
• External validation: COCO_MP set (200 COCO “other” skin images, 132 monkeypox cases).

Data sources include Nafisa et al.'s Monkeypox Skin Lesion Dataset, additional images from Sitaula & Shahi, various clinical reports, and controlled samples such as 11k Hands and FFHQ for healthy controls. Annotation involved two infectious-disease physicians. Augmentations comprise geometric transformations (rotation, translation, flipping), Gaussian noise injection, measured adjustments of brightness and contrast, and oversampling to address class imbalance (Yang et al., 2022).

Training uses the Adam optimizer (initial learning rate 0.001, momentum 0.99), batch size 48, L2 and L1 regularization (as specified above), and an exponential decay learning-rate schedule. Early stopping is based on validation loss (Yang et al., 2022).

4. Performance Benchmarks and Ablation Analysis

AICOM-MP achieves state-of-the-art accuracy under multiple evaluation protocols:

• On the Ali et al. public dataset: AICOM-MP achieves 100% binary monkeypox-vs-other accuracy (prior models: 74–90%).
• On the experimental dataset (6124 images):
  • Precision = 0.9650
  • Recall = 0.9634
  • F₁ score = 0.9635

Ablation experiments on the COCO_MP benchmark show the contribution of each architectural component:

Pipeline Variant	Accuracy (%)
No vision layers	40.66
Classification only	88.55
+ Restoration unit	91.57
+ Background removal (Level-1)	95.18
+ Skin segmentation (Level-2)	95.48
+ Both Level-1 & Level-2	96.99

These improvements, ranging from 2–8 percentage points per stage, are consistent across data splits and ablations. Formal statistical tests were not reported (Yang et al., 2022).

5. Resource-Constrained Deployment: System Design and Optimization

AICOM-MP is engineered for use in computationally limited environments. Model size (EfficientNet-B7 + U²-Net + FCNResNet10) is approximately 150 MB. Inference is CPU-only by default, peaking at about 2 GB RAM and 50% CPU on a quad-core processor. No GPU is required, but optional acceleration is supported for AMC AI edge boxes (e.g., mobile NPUs). Web service response times average 300–600 ms per image on a 4-vCPU/8 GB RAM cloud VM.

Offline deployment is realized via ONNX/TensorFlow Lite wrappers, allowing AMC boxes or compatible mobile devices to conduct inference without an internet connection. If connectivity is intermittent, results are queued locally and synced upon reconnection. This hardware-agnostic approach ensures operability even in field conditions with unreliable power and connectivity (Yang et al., 2022).

6. Integration, Open Source, and Methodological Framework

AICOM-MP is released via GitHub (github.com/Tim-Yang-YTY/AICOM) with the following resources:

• Dataset and all relevant splits/annotation files.
• Modular Python source code (training scripts, model definitions, FastAPI/Flask endpoints).
• Sample client code (Python, JavaScript).
• Docker images for rapid deployment.

Integration into AMCs follows a documented process: retrieving the Docker image, deploying the container, exposing the appropriate port (/predict endpoint), and configuring the AMC gateway to forward requests. ONNX-bundled models are provided for purely offline scenarios. Continuous integration and model updating pipelines are maintained, allowing operators to lock to particular model/dataset versions (e.g., v1.0, v1.1). AICOM-MP exemplifies the best practices for robust dataset design, modular attention-pipeline architecture, compute optimization for resource-limited hardware, and continuous update methodology suitable for field epidemiology (Yang et al., 2022).

7. Generalization and Ethical Considerations

The methodological framework developed through AICOM-MP is generalized for other priority diseases relevant to resource-constrained environments. The modular pipeline supports rapid adaptation: by retraining or swapping classifier heads, the same segmentation/localization strategy can be applied to other diseases (tuberculosis, malaria, diabetic retinopathy). Ethical guidelines include performance monitoring across demographics, open data and code for peer review, and adherence to continuous improvement and transparency standards.

AICOM-MP demonstrates how AI-driven diagnostic tools can be open-sourced, rigorously evaluated, and containerized for integration into next-generation mobile health delivery systems, with explicit attention to resource constraints and equitable global access (Yang et al., 2022).