bidhunb/cataract_detection

Cataract AI Diagnostic System

A deep learning-based application for detecting cataracts using both Fundus and Slit-Lamp eye images.

Features

• —Dual-Model Architecture:
• —Fundus Analysis: Binary classification (Normal vs. Cataract).
• —Slit-Lamp Analysis: Multi-class classification (Normal, Immature, Mature).
• —Preprocessing Pipeline:
• —Green Channel Extraction
• —Denoising (Gaussian Blur)
• —Contrast Enhancement (CLAHE)
• —Interactive Web UI: Built with Flask, featuring drag-and-drop uploads and visualization of preprocessing steps.
• —Deep Learning: Powered by DenseNet169.

Directory Structure

project_root/
├── backend/                   # All Python Code
│   ├── app/                   # Flask Application
│   ├── ml/                    # Machine Learning Logic (Models, Preprocessing, Training)
│   └── config.py              # Configuration
├── data/                      # Datasets
│   ├── raw/                   # Original Datasets (Fundus Binary, Multiclass, Slit-Lamp)
│   └── processed/             # Processed images
├── saved_models/              # Trained model weights
├── run_app.py                 # Application Entry Point
└── README.md                  # Project Information

Installation

1. 1.Clone the repository (if applicable).
2. 2.Install Dependencies:

   pip install torch torchvision numpy opencv-python Flask Pillow tqdm sklearn

Note: Ensure you have a CUDA-enabled GPU for faster training, though CPU is supported.

Usage

1. Running the Application

Start the Flask web server:

python run_app.py

Open your browser and navigate to: http://127.0.0.1:5000

2. Training the Models

The application requires trained models to function effectively.

Train Fundus Binary Model:

python backend/ml/training/train_binary.py --epochs 20

Train Fundus Multiclass Model:

python backend/ml/training/train_multiclass.py --epochs 20

Train Slit-Lamp Model:

python backend/ml/training/train_slit_lamp.py --epochs 20

Models are saved to the `savedmodels/` directory._

How it Works

1. 1.Upload: Select a Fundus or Slit-Lamp image in the respective section.
2. 2.Analysis: The image goes through the preprocessing pipeline.
3. 3.Inference: The processed image is passed to the specific DenseNet169 model.
4. 4.Result: The UI displays the original image, preprocessing steps, diagnosis, and confidence score.

Configuration

Adjust hyperparameters and paths in backend/config.py.

for processing commands

How to Run:

Fundus Binary Only:
python backend/tools/prepare_datasets.py --dataset binary
Fundus Multiclass Only:
bash
python backend/tools/prepare_datasets.py --dataset multiclass
Slit Lamp Only:
bash
python backend/tools/prepare_datasets.py --dataset slit_lamp
All Datasets (Default):
bash
python backend/tools/prepare_datasets.py --dataset all
Disable Augmentation (only split/preprocess):
bash
python backend/tools/prepare_datasets.py --dataset binary --no-augment

slitlamp flow

1. 1.Loading -- Image is loaded as RGB.
2. 2.Preprocessing (Standardization) -- The pipeline.py applies the following steps specifically for slitlamp (when usegreen_channel=False):

-- Resize: Resized to 224x224 pixels. -- Noise Reduction: Applies Gaussian Blur with a kernel size of (5, 5) to reduce high-frequency noise. -- Color-Aware Enhancement (CLAHE): -- Converts image from RGB to LAB color space. -- Extracts the L (Lightness) channel. -- Applies CLAHE (Contrast Limited Adaptive Histogram Equalization) to the L channel only (clipLimit=2.0, tileGridSize=(8,8)). -- Merges channels back and converts to RGB. Rationale: This enhances contrast without distorting the color information. -- Normalization: Pixel values are scaled to the range [0, 1] (divided by 255.0). 3. Splitting -- The dataset is split into Train, Validation, and Test sets (default: 70% / 15% / 15%). 4. Offline Augmentation (Train Set Only) -- If augmentation is enabled, the definitions from augmentations.py (gettraintransforms(imagetype='slitlamp')) are applied to generate 4 additional copies per image:

-- Random Horizontal Flip: 50% probability. -- Random Rotation: +/- 15 degrees (limited to keep upright orientation). -- Color Jitter: Randomizes Brightness (0.8x-1.2x) and Contrast (0.8x-1.2x). -- Random Affine: Scaling: Zoom in/out by +/- 10% (0.9x to 1.1x). Translation: Shift by up to 5% vertically/horizontally.

training commands

# Fundus Binary
python backend/ml/training/train_binary.py --dry-run
# Train for specific epochs (default is 50)
python backend/ml/training/train_binary.py --epochs 100

# Fundus Multiclass
python backend/ml/training/train_multiclass.py --dry-run
python backend/ml/training/train_multiclass.py --epochs 100

# Slit Lamp
python backend/ml/training/train_slit_lamp.py --dry-run
python backend/ml/training/train_slit_lamp.py --epochs 100

======= title: Cataract Detection emoji: 🏢 colorFrom: gray colorTo: pink sdk: docker pinned: false license: mit short_description: AI-Based Cataract Detection Using Deep Learning Models

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference c48679b60b88a8824a32eaa95fa6c1bb335842ad

Run python -m unittest backend/tests/test_risk_assessment.py to ensure backend logic still evaluates correctly and tests pass.