Limitations

Medical Disclaimer: This application is for educational and research purposes only. It is NOT a substitute for professional medical advice, diagnosis, or treatment.

Medical usage limitations

Not a diagnostic tool

This model has critical limitations that prevent its use as a clinical diagnostic device:

Not FDA approved or medically certified

No regulatory approval for clinical use
Not validated through clinical trials
Does not meet medical device standards (FDA, CE, etc.)
Should never replace professional dermatological examination

Training data limitations

Training dataset may not represent all populations equally
Potential bias in skin tone representation
Limited coverage of rare lesion variants
Dataset age and source not verified in production

No liability for medical decisions

Predictions should not influence medical treatment
Always consult qualified healthcare professionals
Developers assume no liability for medical outcomes
Tool may produce false positives or false negatives

Required disclaimers

Any deployment of this application must prominently display:

Required User Warnings:

This tool is for educational purposes only
Not intended for medical diagnosis or treatment decisions
Consult a dermatologist for any skin concerns
Do not delay seeking medical care based on results
Accuracy is not guaranteed and varies by image quality

Technical limitations

Model accuracy constraints

The model’s predictive capability has inherent boundaries: Accuracy factors:

Image quality: Blurry, poorly lit, or low-resolution images reduce accuracy
Lesion positioning: Partial lesions or multiple lesions in frame affect results
Skin tone: Model performance may vary across different skin tones
Image artifacts: Filters, compression, or editing can distort predictions
Context missing: No patient history, lesion evolution, or clinical context

The model outputs a confidence percentage, but high confidence does not guarantee correctness. A 95% confidence prediction may still be wrong.

Classification limitations

The model is restricted to 7 specific lesion types:

Actinic Keratoses
Basal Cell Carcinoma
Benign Keratoses
Dermatofibroma
Melanoma
Melanocytic Nevus
Vascular Lesion

What the model cannot detect:

Lesion types not in the training set
Normal skin (will force-classify into one of 7 categories)
Non-skin objects (will still produce a prediction)
Multiple lesions simultaneously (processes whole image)

The model uses softmax output, meaning it always assigns probabilities summing to 100% across the 7 classes, even for completely irrelevant images.

Input constraints

Image requirements:

Format: JPEG, PNG, WebP, or other browser-supported formats
Size: Any size (automatically resized to 75×100)
Color: RGB color images (grayscale may produce unpredictable results)
Content: Single lesion centered in frame for best results

Unsupported inputs:

Video files (only static images)
Dermoscopic images (model trained on standard photography)
Infrared or specialized medical imaging
Multi-spectral imaging

Image preprocessing artifacts

The model resizes all inputs to 75×100 pixels using nearest neighbor interpolation:

tensorImg = tf.browser.fromPixels(imgtag)
            .resizeNearestNeighbor([75, 100])
            .toFloat().expandDims();

Implications:

High-resolution details are lost during downsampling
Aspect ratio distortion if original image is not 3:4 ratio
Nearest neighbor interpolation may introduce aliasing
Very small lesions (less than 10% of image) may lose critical features

Browser and platform limitations

Compatibility constraints

Browser requirements

Modern browsers only (Chrome 57+, Firefox 52+, Safari 11+, Edge 79+)

JavaScript required

Application fails completely if JavaScript is disabled

WebGL recommended

Significantly slower without WebGL support (10x slower on CPU)

Memory requirements

May crash on devices with less than 2 GB RAM

Mobile device constraints

Performance issues:

Slower inference (400-1000 ms vs 50-150 ms on desktop)
Battery drain during repeated predictions
Large initial download (99 MB) on cellular connections
May not work on older mobile browsers

Camera integration:

Current implementation uses file upload only
No real-time camera feed analysis
No multi-shot or burst mode support

Network dependencies

First-time load requires internet connection:

99.3 MB model download required
Model won’t load on slow/unstable connections
Users on metered connections incur data charges
Offline use only possible after initial cached load

Privacy and security limitations

Data handling

Current implementation processes images entirely client-side: Privacy positives:

Images never uploaded to server
No image storage or logging
Predictions computed locally
HIPAA-compliant processing (no data transmission)

Privacy concerns:

No audit trail or prediction logging
Cannot review past predictions
No data for model improvement or debugging
Browser history/cache may retain images locally

Security considerations

Client-side vulnerabilities

Browser-based execution introduces security risks:

Model tampering: Malicious actors could modify model.json or weights
XSS attacks: Vulnerable if application doesn’t sanitize inputs
Supply chain: Dependency on TensorFlow.js CDN integrity
No authentication: Anyone can access and use the model

Mitigation strategies:

Serve model files over HTTPS only
Implement Subresource Integrity (SRI) for TensorFlow.js
Use Content Security Policy (CSP) headers
Regular security audits of dependencies

No user authentication or rate limiting

Current implementation lacks access controls:

Anyone can use the model unlimited times
No abuse prevention or rate limiting
No usage analytics or monitoring
Cannot block malicious users

For production deployments, consider adding authentication and usage tracking.

Image quality limitations

Optimal image characteristics

For best results, images should meet these criteria:

Characteristic	Optimal	Acceptable	Poor
Lighting	Bright, even, natural light	Indoor lighting	Dark, shadowy, extreme glare
Focus	Sharp, clear lesion edges	Slight blur	Out of focus, motion blur
Distance	Lesion fills 30-70% of frame	10-90% of frame	Less than 10% or more than 95% of frame
Angle	Perpendicular to skin surface	Up to 30° angle	Extreme angles, side views
Background	Plain skin surface	Some hair/skin texture	Clothing, objects in frame
Resolution	More than 500px on shortest side	200-500px	Less than 200px

The model was trained on dermatological images with specific characteristics. Real-world smartphone photos may differ significantly in quality and composition.

Common image problems

Poor lighting reduces accuracy

Too dark: Cannot distinguish lesion features
Too bright: Washed out colors and lost texture
Shadows: Creates false boundaries and color variations
Flash reflection: Specular highlights obscure lesion

Recommendation: Use natural daylight or soft, diffuse indoor lighting

Compression artifacts introduce noise

Heavy JPEG compression can distort lesion characteristics:

Blocky artifacts alter texture perception
Color banding affects pigmentation analysis
Edge distortion impacts border irregularity detection

Recommendation: Use high-quality JPEG (90+) or PNG format

Scaling and aspect ratio distortion

The model’s 75×100 input size creates aspect ratio constraints:

Images are stretched/squished to 3:4 aspect ratio
Circular lesions may appear elliptical
Affects shape-based features in classification

Recommendation: Crop images to ~3:4 aspect ratio before upload

Functional limitations

Missing features

The current implementation lacks several important capabilities: No batch processing:

Can only analyze one image at a time
No comparison between multiple images
Cannot track lesion changes over time

No result history:

Predictions disappear on page reload
Cannot review or export past results
No tracking of prediction confidence trends

No additional context:

Cannot input patient age, sex, or history
No lesion size measurement
No anatomical location consideration
No symptom information (itching, bleeding, etc.)

Limited output information:

Only provides class name and confidence
No explanation of features detected
No visualization of what model “sees”
No alternative diagnoses or differential considerations

Error handling limitations

Basic error handling may not cover all edge cases:

try {
    if(imgtag.src == ""){
        alert("Select an Image to Classify")
        return
    }
    // ... prediction code
} catch(error) {
    alert("Error Classifying Image")
}

Unhandled scenarios:

Model load failure (shows generic error)
Out of memory errors on low-end devices
Corrupted or malformed image files
Very large images causing browser freeze
Network timeout during model download

Ethical and bias limitations

Algorithmic bias concerns

Machine learning models can perpetuate biases present in training data.

Potential bias sources:

Skin tone: Model may perform better on lighter skin tones if training data is imbalanced
Demographics: Age, sex, and ethnicity representation in training set unknown
Geographic: Lesions more common in certain populations may be underrepresented
Image source: Medical vs. consumer photos may have different characteristics

Transparency limitations:

Training dataset composition not documented in codebase
No per-demographic accuracy metrics available
Cannot verify fairness across protected groups

Responsible AI considerations

Deployment of this model requires careful consideration:

Informed consent and user education

Accessibility limitations

Current implementation may exclude users:

No screen reader support for predictions
Color-dependent UI may not work for colorblind users
No keyboard navigation support
Requires manual file upload (no voice input)
English-only interface and class names

Regulatory and legal limitations

Legal Disclaimer: Using this tool for medical diagnosis may violate healthcare regulations in many jurisdictions.

Compliance requirements for medical use

To legally deploy as a medical device, would require:

FDA 510(k) clearance or De Novo classification (USA)
CE marking under MDR (European Union)
Clinical validation studies
Quality management system (ISO 13485)
Post-market surveillance and adverse event reporting
Labeling and instructions for use

Current status: None of the above are implemented

Liability and risk

Developers and deployers may face legal liability if:

Application is marketed as a diagnostic tool
Users delay medical care based on predictions
Misdiagnosis leads to patient harm
Privacy breaches occur (even if client-side)
Regulatory requirements are not met

Future improvements

Recognized limitations that could be addressed in future versions:

Model validation: Comprehensive accuracy testing across demographics
Explainability: Grad-CAM or attention maps showing decision factors
Uncertainty quantification: Bayesian approaches to estimate prediction confidence
Expanded classes: Support for more lesion types and “unknown” category
Quality checks: Automatic image quality assessment before prediction
Temporal analysis: Track lesion changes over multiple images
Accessibility: WCAG 2.1 AA compliance for interface
Mobile optimization: Native apps with camera integration
Offline-first: Service worker for full offline capability
Federated learning: Privacy-preserving model improvements from user data

Contributions and improvements are welcome. See the project repository for development guidelines.

Get Started

Understanding the Model

Using the API

Technical Details

Resources

Medical usage limitations

Not a diagnostic tool

Required disclaimers

Technical limitations

Model accuracy constraints

Classification limitations

Input constraints

Browser and platform limitations

Compatibility constraints

Browser requirements

JavaScript required

WebGL recommended

Memory requirements

Mobile device constraints

Network dependencies

Privacy and security limitations

Data handling

Security considerations

Image quality limitations

Optimal image characteristics

Common image problems

Functional limitations

Missing features

Error handling limitations

Ethical and bias limitations

Algorithmic bias concerns

Responsible AI considerations

Regulatory and legal limitations

Compliance requirements for medical use

Liability and risk

Future improvements

Build docs developers (and LLMs) love

Get Started

Understanding the Model

Using the API

Technical Details

Resources

​Medical usage limitations

​Not a diagnostic tool

​Required disclaimers

​Technical limitations

​Model accuracy constraints

​Classification limitations

​Input constraints

​Browser and platform limitations

​Compatibility constraints

Browser requirements

JavaScript required

WebGL recommended

Memory requirements

​Mobile device constraints

​Network dependencies

​Privacy and security limitations

​Data handling

​Security considerations

​Image quality limitations

​Optimal image characteristics

​Common image problems

​Functional limitations

​Missing features

​Error handling limitations

​Ethical and bias limitations

​Algorithmic bias concerns

​Responsible AI considerations

​Regulatory and legal limitations

​Compliance requirements for medical use

​Liability and risk

​Future improvements

Build docs developers (and LLMs) love

Medical usage limitations

Not a diagnostic tool

Required disclaimers

Technical limitations

Model accuracy constraints

Classification limitations

Input constraints

Browser and platform limitations

Compatibility constraints

Mobile device constraints

Network dependencies

Privacy and security limitations

Data handling

Security considerations

Image quality limitations

Optimal image characteristics

Common image problems

Functional limitations

Missing features

Error handling limitations

Ethical and bias limitations

Algorithmic bias concerns

Responsible AI considerations

Regulatory and legal limitations

Compliance requirements for medical use

Liability and risk

Future improvements