How jayeshdambale/retinal-vessel-segmentation works
Project Type and Purpose
• This is a deep-learning image-segmentation pipeline (“tool”) specialized for retinal vessel detection using a U-Net architecture.
• It ingests raw fundus images, prepares datasets (DRIVE, CHASE), trains a U-Net model, saves weights, and runs inference to output segmented vessel masks.
Main Components
a. Data Ingestion & Preparation
• Raw dataset files (e.g., DRIVE)
• Script: prepare_datasets_DRIVE.py
• Output: preprocessed images & masks in a “data/” staging area
b. Training Pipeline
• Script: run_training.py
• Core logic (imported from src/): model instantiation (U-Net encoder/decoder), loss definition, optimizer, training loop
• Dependencies: TensorFlow/Keras, NumPy, OpenCV, Matplotlib
• Output: trained model weights in models/
c. Testing & Inference
• Script: run_testing.py
• Loads saved weights, runs prediction on new images, computes metrics, saves segmented outputs to an output directory
d. Configuration
• configuration.txt holds hyperparameters (learning rate, batch size, dataset paths)
e. Utilities
• Any helper functions (e.g., IO routines, metric calculations) in utils/
Relationships & Data Flow
1. “Raw Images” → prepare_datasets_DRIVE.py → “Preprocessed Data”
2. Preprocessed Data → run_training.py → “Trained Weights”
– Internally, U-Net (Encoder ↔ Decoder) transforms tensor batches → loss → backprop
3. New Images + Trained Weights → run_testing.py → “Segmented Masks + Metrics”
Architectural Patterns & Design Principles
• Pipeline pattern: clear separation between data prep, training, inference
• Modular code: scripts invoke functions in src/ (model code), utils/ (helpers)
• Config-driven: hyperparameters and paths in configuration.txt
• Reproducibility: gitignore, requirements.txt, consistent folder layout