top of page
Search

From Science Fair to Real-World Impact: The Future of NOMA AI

  • Writer: Anie Etor-Udofia
    Anie Etor-Udofia
  • Apr 1
  • 2 min read

The Journey So Far

NOMA AI began as a question: Could a $200 Raspberry Pi device provide preliminary skin cancer screening? After nearly a year of development, the answer is yes.


What We Built:

  • A 24-class skin condition classifier (4 malignant, 20 benign, 1 normal)

  • A clinical assessment wizard based on ABCDE criteria

  • A TFLite-optimized model running on Raspberry Pi 4

  • A touchscreen GUI with swipe navigation

  • Grad-CAM heatmaps for interpretability

  • LED feedback for immediate risk communication

  • Health passport for longitudinal tracking

Performance Metrics:

  • Validation Accuracy: 61.5%

  • Melanoma F1-Score: 0.933

  • Normal Skin F1-Score: 0.930

  • Inference Time: 0.3-0.5 seconds on Pi 4


Lessons Learned

1. Class Imbalance is Real - Malignant classes had fewer images, but they needed the most attention. Class weights solved this, but future iterations need more data for rare cancers.

2. Explainability is Non-Negotiable - Users don't trust a "black box" diagnosis. Grad-CAM heatmaps and top-3 alternative diagnoses built trust during testing.

3. Touch Interfaces Need Large Targets - Buttons need at least 80px height for fat-finger friendly operation. Swipe gestures need a threshold (100px) to avoid accidental navigation.

4. Error Handling Must Be Aggressive - Medical devices can't crash. The app has fallbacks for every component: camera failure (dummy frames), GPIO failure (mock mode), low confidence (retake prompt).

5. The ABCDE Framework is Genius - Clinicians have used ABCDE for decades. Fusing it with AI creates a system that's both accurate and clinically valid.


What's Next?

1. Clinical Validation - Partner with dermatologists to validate against biopsy-confirmed cases. Real-world testing is the next critical step.

2. Expanded Dataset

  • Add more images for underrepresented skin types (Fitzpatrick IV-VI)

  • Include more malignant samples

  • Add dermatoscopic images (currently using standard photos)

3. Model Improvements

  • Try EfficientNet or ConvNeXt architectures

  • Implement test-time augmentation for higher confidence

  • Add ensemble methods (multiple models voting)

4. Hardware Refinements

  • Custom PCB for LEDs and buttons

  • Battery management for longer runtime

  • 3D-printed enclosure (currently acrylic sheets)

5. Regulatory Pathway

  • FDA Class I (general controls) for non-diagnostic screening

  • Explore Class II (special controls) for clinical use

  • HIPAA compliance for patient data

6. Open Source ReleaseThe code, dataset, and hardware designs will be open-sourced for:

  • Low-resource clinics

  • Research institutions

  • Other developers building medical AI


Call to Action

NOMA AI is proof that accessible healthcare technology doesn't require millions in funding. A $200 device, open-source tools, and dedicated effort can create something that makes a difference.

If you're interested in:

  • Contributing: GitHub issues and pull requests welcome

  • Clinical Collaboration: Dermatologists, please reach out

  • Deployment: Let's discuss pilot programs

  • Research: The dataset is available on Kaggle


The Bigger Picture

Skin cancer doesn't wait. Neither should screening. NOMA AI puts preliminary evaluation in anyone's hands—not to replace doctors, but to empower people with information before it's too late.

This project started as a science fair entry, but it has the potential to become something more. With validation, funding, and collaboration, NOMA AI could help catch skin cancer earlier, especially in communities with limited dermatology access.

The future of healthcare is accessible, explainable, and patient-centered. NOMA AI is a step in that direction.

Comments

bottom of page