Skip to content

abhay-dev2901/Carbs-ai

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
data
data
 
 
frontend
frontend
 
 
models
models
 
 
scripts
scripts
 
 
src
src
 
 
.gitignore
.gitignore
 
 
README.md
README.md
 
 
app.py
app.py
 
 
requirements.txt
requirements.txt
 
 

Repository files navigation

🍽️ Type 1 Diabetes Carb Estimation System

An AI/ML-powered system for gram-precise net carbohydrate estimation and glycemic load prediction from meal photos, designed to support safer insulin dosing decisions for Type 1 diabetics.

Python TensorFlow License

📋 Table of Contents

🎯 Problem Statement

The Challenge

Type 1 diabetes requires precise insulin dosing based on carbohydrate intake. The standard rule is:

  • 1 unit of insulin per 10-15g of carbohydrates

However, manual carbohydrate counting has significant limitations:

  • ±20-30% error rate in typical manual counting
  • Each 5g error can cause 25-50 mg/dL blood glucose fluctuations
  • Poor estimation increases risks of:
    • Hypoglycemia (low blood sugar)
    • Hyperglycemia (high blood sugar)
    • Long-term diabetic complications

Our Solution

This system uses computer vision and deep learning to:

  • Identify Indian dishes from meal photos
  • Estimate net carbohydrates with ±10g accuracy (v1 target)
  • Calculate glycemic load for better blood sugar management
  • Provide insulin dosing recommendations as a decision support tool

✨ Key Features

🎯 Core Capabilities

  • Dish Classification: Identifies 6 Indian dishes with 89% accuracy

    • Biryani, Dal, Halwa, Poha, Rasgulla, Roti

  • Carbohydrate Estimation: Predicts net carbs with 82% accuracy within ±10g

    • Mean Absolute Error: 5.92g
    • Median Error: 3.05g

  • Glycemic Load Calculation:

    • Computes GL from GI and net carbs
    • Classifies as Low/Medium/High GL
    • Provides blood sugar impact warnings

  • Insulin Recommendations:

    • Customizable insulin-to-carb ratios (default: 1 unit per 12g)
    • Individual ratio support (10-15g per unit)

  • Clinical Context:

    • Blood glucose impact estimates
    • Safety warnings and disclaimers
    • Comparison to manual counting accuracy

🔧 Technical Features

  • Enhanced Data Augmentation: Rotation, shear, zoom, brightness adjustments
  • Robust Training: Early stopping, learning rate scheduling, model checkpointing
  • Comprehensive Evaluation: MAE, RMSE, R², error distribution, per-class analysis
  • Production-Ready: Modular code structure, error handling, validation

📊 Performance Metrics

Current Performance (v1)

Classifier Performance

  • Accuracy: 89.0%
  • Top-K Accuracy: 98.5%
  • Per-Class Performance:
    • Biryani: 90% precision, 90% recall
    • Dal: 91% precision, 100% recall
    • Halwa: 100% precision, 100% recall
    • Poha: 89% precision, 80% recall
    • Rasgulla: 100% precision, 100% recall
    • Roti: 100% precision, 90% recall

Regressor Performance

  • Mean Absolute Error (MAE): 5.92g ✅ (Target: <10g)
  • Root Mean Squared Error (RMSE): 8.97g
  • R² Score: 0.77
  • Mean Absolute Percentage Error (MAPE): 18.61%

Accuracy Thresholds

  • Within ±5g: 58.5%
  • Within ±10g: 82.0% ✅ (Target: >80%)
  • Within ±15g: 89.5%

Error Distribution

  • Min Error: 0.03g
  • Max Error: 29.03g
  • Median Error: 3.05g
  • 75th Percentile: 8.00g
  • 95th Percentile: 22.52g

Per-Class Carb Prediction

Dish True Carbs Predicted Error MAE
Biryani 45.0g 39.9g 7.4g 7.4g
Dal 12.0g 13.3g 2.7g 2.7g
Halwa 60.0g 52.4g 10.6g 10.6g
Poha 28.0g 30.1g 4.0g 4.0g
Rasgulla 35.0g 38.4g 7.5g 7.5g
Roti 18.0g 19.2g 3.5g 3.5g

Clinical Significance

  • Target Accuracy: ±10g for safe insulin dosing ✅ ACHIEVED
  • Blood Glucose Impact: Average error (5.9g) ≈ 30-59 mg/dL change
  • Comparison: Manual counting typically has ±20-30% errors
  • Status: Model meets v1 target accuracy threshold

📦 Dataset

Overview

  • Total Images: 1,000 images
  • Classes: 6 Indian dishes
  • Split: 80% training (800 images), 20% validation (200 images)
  • Source: Custom dataset + Indian Food Images dataset

Distribution

Dish Images Status
Biryani 150 ⚠️ Need 50+
Dal 250 ✅ Good
Halwa 250 ✅ Good
Poha 150 ⚠️ Need 50+
Rasgulla 50 ⚠️ Need 150+
Roti 150 ⚠️ Need 50+

Nutrition Data

Each dish has associated nutrition information:

Dish Net Carbs (g) Glycemic Index
Biryani 45 70
Dal 12 30
Halwa 60 75
Poha 28 65
Rasgulla 35 65
Roti 18 55

Data Quality

  • Diversity: Various lighting, angles, backgrounds, portion sizes
  • Validation: All images validated for readability and quality
  • Augmentation: Enhanced augmentation during training

🏗️ Model Architecture

Classifier Architecture

Input (224×224×3)
    ↓
MobileNetV2 (ImageNet pretrained, frozen)
    ↓
Global Average Pooling (GAP)
    ↓
Batch Normalization
    ↓
Dense(256) + ReLU + Dropout(0.5)
    ↓
Dense(128) + ReLU + Dropout(0.3)
    ↓
Dense(6) + Softmax
    ↓
Output: Dish Classification

Training Details:

  • Optimizer: Adam (lr=0.001)
  • Loss: Categorical Crossentropy
  • Callbacks: Early Stopping, ReduceLROnPlateau, ModelCheckpoint
  • Epochs: 30 (with early stopping)
  • Data Augmentation: Rotation, shear, zoom, brightness, horizontal flip

Regressor Architecture

Input (224×224×3)
    ↓
Frozen Classifier Feature Extractor (GAP layer)
    ↓
Batch Normalization
    ↓
Dense(256) + ReLU + Dropout(0.4)
    ↓
Dense(128) + ReLU + Dropout(0.3)
    ↓
Dense(64) + ReLU + Dropout(0.2)
    ↓
Dense(1) + Linear
    ↓
Output: Net Carbohydrates (grams)

Training Details:

  • Optimizer: Adam (lr=0.001)
  • Loss: Huber Loss (robust to outliers)
  • Metrics: MAE, MSE
  • Callbacks: Early Stopping, ReduceLROnPlateau, ModelCheckpoint
  • Epochs: 50 (with early stopping)

Key Design Decisions

  1. Transfer Learning: MobileNetV2 pretrained on ImageNet for feature extraction
  2. Frozen Base: Prevents overfitting on small dataset
  3. Huber Loss: More robust to outliers than MSE
  4. Progressive Dropout: Higher dropout in deeper layers
  5. Feature Reuse: Regressor uses classifier features for efficiency

🚀 Installation

Prerequisites

  • Python 3.9 or higher
  • pip package manager
  • 4GB+ RAM (8GB recommended)
  • GPU optional but recommended for faster training

Step 1: Clone the Repository

git clone https://github.com/yourusername/Carbs-ai.git
cd Carbs-ai

Step 2: Install Dependencies

pip install -r requirements.txt

Required Packages:

tensorflow==2.15.0
pandas==2.2.2
numpy==1.26.4
Pillow==10.4.0
scikit-learn==1.5.1

Step 3: Verify Installation

python -c "import tensorflow as tf; print(f'TensorFlow {tf.__version__}')"
python -c "import pandas as pd; print(f'Pandas {pd.__version__}')"

🏃 Quick Start

1. Train Models

Train Classifier (takes ~5-10 minutes):

python src/train_classifier.py

Train Regressor (takes ~5-10 minutes):

python src/train_regressor.py

2. Make a Prediction

python src/predict.py data/images/biryani/0d81432b55.jpg

Output Example:

======================================================================
TYPE 1 DIABETES CARB ESTIMATION - PREDICTION RESULTS
======================================================================

🍽️  Identified Dish: BIRYANI
   Confidence: 99.0%

📊 Net Carbohydrates: 30.9g
📈 Glycemic Index (GI): 70
📉 Glycemic Load (GL): 21.7
   High GL — Rapid blood sugar rise ⚠️

💉 Insulin Recommendation:
   Suggested dose: 2.6 units
   (Based on 1 unit per 12g carbs)

3. Evaluate Models

python src/evaluate.py

📖 Usage Guide

Making Predictions

Basic Prediction

python src/predict.py <image_path>

Custom Insulin Ratio

python src/predict.py <image_path> --insulin-ratio 10

(Default is 12g per unit; adjust for individual needs)

Example Predictions

# Biryani
python src/predict.py data/images/biryani/0d81432b55.jpg

# Dal
python src/predict.py data/images/dal/006aeeea8d.jpg

# Rasgulla
python src/predict.py data/images/rasgulla/<image_name>.jpg

Training Models

Classifier Training

python src/train_classifier.py

What it does:

  • Loads images from data/images/
  • Applies data augmentation
  • Trains MobileNetV2-based classifier
  • Saves model to models/classifier.h5
  • Shows validation accuracy

Regressor Training

python src/train_regressor.py

What it does:

  • Loads trained classifier
  • Extracts features from GAP layer
  • Trains regression head for carb prediction
  • Saves model to models/regressor.h5
  • Shows comprehensive evaluation metrics

Evaluating Models

python src/evaluate.py

Provides:

  • Classification accuracy and confusion matrix
  • Regression metrics (MAE, RMSE, R², MAPE)
  • Accuracy within thresholds (±5g, ±10g, ±15g)
  • Error distribution statistics
  • Per-class performance analysis
  • Clinical significance assessment

Expanding Dataset

See DATASET_GUIDE.md for detailed instructions on:

  • Finding compatible datasets
  • Downloading from Kaggle
  • Organizing new images
  • Validating dataset quality

Quick expansion:

# Organize images from downloaded dataset
python scripts/expand_dataset.py <source_folder> --target data/images

# Check current dataset status
python scripts/expand_dataset.py data/images --count-only

📁 Project Structure

Carbs-ai/
├── data/
│   ├── images/              # Training images organized by dish
│   │   ├── biryani/
│   │   ├── dal/
│   │   ├── halwa/
│   │   ├── poha/
│   │   ├── rasgulla/
│   │   └── roti/
│   ├── nutrition.csv        # Net carbs and GI data
│   └── raw/                 # Raw downloaded datasets
│
├── models/                  # Trained model files
│   ├── classifier.h5
│   ├── classifier_best.h5
│   ├── regressor.h5
│   └── regressor_best.h5
│
├── src/
│   ├── train_classifier.py # Classifier training script
│   ├── train_regressor.py  # Regressor training script
│   ├── predict.py          # Prediction script
│   ├── evaluate.py        # Model evaluation script
│   └── utils/
│       ├── __init__.py
│       └── diabetic_check.py  # Utility functions
│
├── scripts/
│   └── expand_dataset.py   # Dataset expansion utility
│
├── requirements.txt         # Python dependencies
├── README.md               # This file
└── DATASET_GUIDE.md        # Dataset expansion guide

🏥 Clinical Context

Why Accuracy Matters

Insulin Dosing Rule:

  • 1 unit of insulin per 10-15g carbohydrates (individual varies)

Impact of Errors:

  • 5g error → 25-50 mg/dL blood glucose change
  • 10g error → 50-100 mg/dL blood glucose change
  • 15g error → 75-150 mg/dL blood glucose change

Current System Performance:

  • Average error: 5.9g (30-59 mg/dL impact)
  • 82% of predictions within ±10g
  • Better than manual counting (±20-30% errors)

Glycemic Load (GL) Classification

  • Low GL (≤10): Gradual blood sugar rise ✅ Safe
  • Medium GL (11-19): Moderate blood sugar rise ⚠️ Caution
  • High GL (≥20): Rapid blood sugar rise ⚠️ High caution

Formula: GL = (GI × Net Carbs) / 100

Clinical Recommendations

  1. Use as Decision Support: Always verify with healthcare provider
  2. Individual Ratios: Adjust insulin ratio based on personal needs
  3. Monitor Blood Sugar: Track actual vs. predicted impact
  4. Continuous Learning: System improves with more data

⚠️ Safety & Disclaimers

Important Medical Disclaimer

⚠️ THIS IS NOT A MEDICAL DEVICE

  • This system is a research/decision support tool only
  • NOT FDA approved for medical use
  • NOT a substitute for professional medical advice
  • Always consult with healthcare providers for insulin dosing
  • Individual insulin ratios vary (10-15g per unit)
  • Blood sugar responses vary by person

Limitations

  1. Limited Dataset: Trained on 6 dishes, 1000 images
  2. Portion Size: Assumes standard serving sizes
  3. Variability: Different preparations may vary in carbs
  4. Accuracy: 82% within ±10g (not 100% accurate)
  5. Clinical Validation: Not clinically validated

Best Practices

  • ✅ Use as a starting point for carb estimation
  • ✅ Verify predictions with nutrition labels when available
  • ✅ Monitor blood sugar after meals
  • ✅ Adjust based on personal experience
  • ❌ Don't rely solely on predictions
  • ❌ Don't use without healthcare provider consultation

🔮 Future Enhancements

Short-term (v1.1 - v1.5)

  • Expand dataset to 200+ images per dish
  • Add more Indian dishes (10+ total)
  • Portion size estimation
  • Multi-dish meal support
  • Real-time mobile app

Medium-term (v2.0)

  • Clinical validation study
  • User feedback loop for continuous improvement
  • Personalization based on user history
  • Integration with CGM (Continuous Glucose Monitor)
  • Meal planning recommendations

Long-term (v3.0+)

  • Multi-cuisine support (beyond Indian food)
  • Real-time video analysis
  • Community dataset contributions
  • FDA approval pathway exploration
  • Integration with insulin pumps

🤝 Contributing

We welcome contributions! Areas where help is needed:

  1. Dataset Expansion: More images, more dishes
  2. Model Improvements: Better architectures, hyperparameter tuning
  3. Code Quality: Bug fixes, documentation, testing
  4. Clinical Validation: Real-world testing and feedback
  5. Feature Development: New features from roadmap

How to Contribute

  1. Fork the repository
  2. Create a feature branch (git checkout -b feature/AmazingFeature)
  3. Commit your changes (git commit -m 'Add some AmazingFeature')
  4. Push to the branch (git push origin feature/AmazingFeature)
  5. Open a Pull Request

Contribution Guidelines

  • Follow PEP 8 style guide
  • Add docstrings to new functions
  • Include tests for new features
  • Update documentation as needed

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

  • Dataset: Indian Food Images dataset from Kaggle
  • Framework: TensorFlow/Keras for deep learning
  • Architecture: MobileNetV2 for efficient feature extraction
  • Community: Open source ML community for tools and inspiration

📞 Contact & Support

📚 Additional Resources

Made with ❤️ for the Type 1 diabetes community

Remember: This tool is for decision support only. Always consult with healthcare providers for medical decisions. Contributor fix test

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

1 fork
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages