About

The Problem

• Skin cancer affects millions worldwide, yet access to dermatological screening remains limited

• Long wait times—up to 12–16 weeks in some regions

• Geographic barriers isolate rural and Indigenous populations (30–40% longer diagnostic intervals)

• High cost of specialist visits means many suspicious lesions go unexamined until it is too late

• Commercial AI systems show performance disparities of up to 34.7% across skin tones

• Most existing solutions are proprietary black boxes—accurate, but impossible to trust

Melanoma caught early has a 99% survival rate; caught late, that drops to 35%.

NOMA AI was built to change this.

Our Solution

NOMA AI is a portable, explainable, AI-powered skin cancer screening system designed to democratize dermatological assessment. Built on a Raspberry Pi 4 with a custom enclosure, it runs entirely offline—no internet, no cloud, no data leaving the device.

Core Features

• AI Visual Analysis – MobileNetV3 deep learning model trained on 26,838 images across 24 skin conditions (malignant, precancerous, benign, inflammatory, and infectious)

• Grad-CAM Explainability – True gradient-based heatmaps that highlight exactly which image regions influenced the model's decision, validated by perturbation tests (37% confidence drop when high-attention regions masked)

• ITA Skin Tone Calibration – Individual Typology Angle (ITA) continuous skin tone measurement with adaptive contrast enhancement (CLAHE) for darker skin tones, reducing bias to a 2.0% accuracy gap

• ABCDE Clinical Wizard – Step-by-step interactive guide through Asymmetry, Border, Color, Diameter, and Evolution—the gold-standard clinical criteria used by dermatologists worldwide

• ORB Longitudinal Tracking – 500-feature fingerprint matching to track lesions over time with 35-match threshold; detects changes in size, color, and border irregularity across scans

• Immediate Physical Feedback – Tri-color LED system (Red / Yellow / Green) for instant risk communication

• Health Passport – Complete longitudinal health record with timestamps, predictions, confidence scores, ITA measurements, and change detection history

• Cross-Modal Integration – Shares data with THORACIS AI (lung screening system) through Operation Oracle unified platform, enabling paraneoplastic syndrome alerts

How It Works

1. Capture

• High-resolution lesion photography using the 16MP Arducam IMX519 camera

• Real-time preview on the 5-inch touchscreen

2. Analyze

• On-device AI inference via TFLite-optimized MobileNetV3

• Processes in under 3 seconds—no internet required

3. Explain

• Grad-CAM heatmaps show why the AI made its prediction

• ITA calibration displays skin tone and bias risk level

• ABCDE wizard provides clinical rationale in plain language

4. Track

• ORB feature matching creates a unique fingerprint for each lesion

• Subsequent scans automatically detect changes in size, color, and border irregularity

5. Act

• LED feedback (Red/Yellow/Green) provides immediate risk communication

• Health Passport stores all assessments for longitudinal monitoring

Technical Architecture

Hardware

• Processor: Raspberry Pi 4 (4GB RAM) – Real-time inference and system control

• Camera: Arducam IMX519 (16MP) – High-resolution autofocus imaging

• Display: 5-inch touchscreen – Interactive GUI with Qt-based navigation

• Feedback: Tri-color LED array – Red / Yellow / Green physical indicators

• Power: Portable power bank (5V/3A+) – Fully mobile, clinic-ready design

• Audio (Cross-Modal): BOYA BY-MS Lavalier microphones – For THORACIS AI integration

• Total Cost: <$300 CAD (NOMA AI alone) | <$800 CAD (Operation Oracle complete)

Software Pipeline

• Image Capture: Picamera2 / OpenCV – High-resolution lesion photography with real-time preview

• Skin Tone Calibration: CIELAB / ITA – Individual Typology Angle calculation with adaptive CLAHE enhancement

• AI Inference: TensorFlow Lite – MobileNetV3 running entirely on-device, <3 seconds per inference

• Feature Extraction: OpenCV – Automated asymmetry, border irregularity, color distribution, diameter estimation

• Explainability: Grad-CAM / TrueGradCAM – Gradient-based attention maps with perturbation validation

• Clinical Assessment: ABCDE Wizard – Step-by-step interactive clinical criteria with patient history

• Longitudinal Tracking: ORB Feature Matching – 500-feature fingerprint with 35-match threshold

• Risk Fusion: Weighted Ensemble – AI confidence (40%) + ABCDE score (40%) + Patient risk (20%)

• Data Storage: SQLite / JSON – Health Passport with timestamped records

• Cross-Modal Sync: Shared Folder (/opt/oracle_share) – Integration with THORACIS AI for paraneoplastic alerts

Model Architecture

Dataset:

• 26,838 images across 24 classes

• Original clinical data: 13,358 images

• MILK10k: 10,480 images (3× weighting for dark skin)

• GAN-generated: 3,000 synthetic dark skin images

Backbone: MobileNetV3 (ImageNet pre-trained)

Custom Classification Head:

• Global Average Pooling

• Dropout (0.3)

• Dense (512, ReLU) + BatchNorm

• Dropout (0.4)

• Dense (256, ReLU)

• Dropout (0.3)

• Softmax (24 classes)

Training Strategy:

• Stage 1: Classifier training (30 epochs, LR = 0.001)

• Stage 2: Fine-tuning last 40 layers (40 epochs, LR = 0.0001)

• Stage 3: Full model training (30 epochs, LR = 0.00001, if needed)

Deployment Optimization:

• Converted to TensorFlow Lite using tf.lite.Optimize.DEFAULT

• Efficient edge deployment on Raspberry Pi CPU

Risk Calculation

NOMA AI combines multiple factors to generate a comprehensive risk assessment:

Components:

• AI Confidence (40%) – Model prediction with confidence score

• ABCDE Score (40%) – Clinical features (Asymmetry, Border, Color, Diameter, Evolution)

• Patient Risk (20%) – Age, skin type (ITA), family history, sunburn history, symptoms

LED Alert System

• 🔴 RED (≥70%) – High risk → Urgent dermatology referral within 1-2 weeks

• 🟡 YELLOW (40–69%) – Moderate risk → Schedule follow-up, monitor monthly

• 🟢 GREEN (<40%) – Low risk → Continue self-monitoring, annual exam

Recent Updates (2026)

The latest version of NOMA AI includes major enhancements based on real-world testing and feedback:

March–June 2026 Updates

• ITA Skin Tone Calibration – Continuous Individual Typology Angle measurement with adaptive CLAHE contrast enhancement for darker skin tones; bias risk level displayed to user

• True Grad-CAM Heatmaps – Access to intermediate tensors generating actual gradient-based attention maps; perturbation-validated (37% confidence drop on masked regions)

• ORB Longitudinal Tracking – 500-feature fingerprint extraction with 35-match threshold; automatic change detection across scans

• ABCDE Clinical Wizard – Step-by-step interactive guide through clinical criteria with plain-language rationale

• Health Passport – Complete assessment history with timestamps, predictions, confidence, ITA, and skin tone

• Top-3 Alternative Diagnoses – Improves interpretability by showing other likely conditions

• Uncertainty Estimation – Flags predictions with >80% uncertainty, prompting image retake

• Clinical Context Expansion – New inputs: itchiness, sudden onset, recurrence, pain, bleeding

• Educational Tip Banner – Random dermatology insights to promote skin health awareness

• Cross-Modal Sync – Integration with THORACIS AI via /opt/oracle_share for paraneoplastic syndrome alerts

• Bias Report Publication – Full bias analysis and code published on Kaggle with 87% dark skin accuracy (2.0% gap)

Performance Metrics

• Overall Accuracy: 86.67% (n=1,200 validation)

• Light Skin (Fitzpatrick I-II): 89.0%

• Medium Skin (Fitzpatrick III-IV): 84.0%

• Dark Skin (Fitzpatrick V-VI): 87.0%

• Bias Gap (Light vs. Dark): 2.0% (vs. 34.7% in commercial systems)

• Melanoma F1: 1.00 (n=15, Wilson 95% CI [0.78, 1.00])

• Inference Speed: <3 seconds on Raspberry Pi CPU

• Cost: <$300 CAD

Explainability Features

NOMA AI provides five layers of explainability—a fundamentally different design philosophy from commercial black-box systems:

• True Grad-CAM – Actual gradient-based attention maps, validated by perturbation tests (37% confidence drop on masked high-attention regions)

• ITA-Based Transparency – Users see their ITA score, skin tone category, and Bias Risk Level (Low/Medium/High/Highest)

• ABCDE Wizard – Step-by-step clinical rationale for each risk level in plain language

• ORB Match Count – Users see exactly how many features matched (e.g., "47/500 features matched")

• Health Passport – Complete history of scans with timestamps, predictions, and changes

This transparency enables informed decision-making: a user with high bias risk can seek clinical confirmation rather than relying solely on the AI.

Project Goals

NOMA AI was designed to:

• Democratize access to preliminary skin cancer screening in underserved communities

• Bridge the gap between AI automation and clinical reasoning through explainability

• Educate users about skin health, warning signs, and the ABCDE criteria

• Provide explainable results that build trust between patients and AI

• Enable longitudinal tracking through a "health passport" for change detection over time

• Reduce algorithmic bias through ITA-based continuous calibration

• Integrate with lung screening (THORACIS AI) for paraneoplastic syndrome detection

Limitations

NOMA AI is a screening tool—not a diagnostic device. It does not replace:

• Professional dermatological examinations

• Biopsy and histopathology

• In-person medical consultation

• Clinical judgment by qualified healthcare providers

Always consult a healthcare professional for medical concerns.

Open Source

NOMA AI is fully open-source. All hardware designs, software, and documentation are available for replication, modification, and improvement.

• GitHub: github.com/havil/noma-ai

• Kaggle Bias Report: kaggle.com/anieetorudofia/noma-ai-bias-report

• Shared Sync Folder: /opt/oracle_share (for cross-modal integration)

Build your own device. Improve the model. Translate the UI. Democratize skin health.

The Team

NOMA AI was developed as a CYSF 2026 project by a Grade 10 student passionate about accessible healthcare technology. The system integrates machine learning, embedded systems, clinical reasoning, and cross-modal integration into a portable, affordable screening solution.

Operation Oracle—NOMA AI + THORACIS AI—represents a complete, connected approach to democratizing early cancer detection.