DRP Segmentation

Diabetic retinopathy screening prototype built with FastAPI. The project combines retinal image analysis, segmentation, and LLM-based report generation to produce interpretable outputs for screening workflows.

• FastAPI backend with vision and NLP endpoints
• Retina image segmentation pipeline for vessel and lesion analysis
• Prompt/template layer for generating structured reports
• Notebooks for experimentation, training, and evaluation

DRP_segmentation/
├── notebooks/ # Exploration, training, and evaluation notebooks
├── src/
│ ├── main.py # FastAPI app entry point
│ ├── routes/ # API routes for vision and NLP
│ ├── controllers/ # Application logic
│ ├── engine/ # Image preprocessing and segmentation
│ ├── llm/ # LLM provider and template utilities
│ ├── helpers/ # Settings and shared helpers
│ ├── models/ # Enums and model-related assets
│ └── requirements.txt # App-specific dependencies
├── requirements.txt # Root dependencies for the app
├── Dockerfile # Containerized runtime
└── README.md

• Analyze a retinal image and return an overlay plus a generated report
• Generate structured NLP outputs from model results
• Load segmentation weights at startup for vessel and lesion inference
• Support configurable LLM providers through environment settings

• Python 3.10+
• A valid Cohere API key if you want to use the LLM-backed report generation
• Trained model weights placed at the paths expected by src/main.py

pip install -r requirements.txt

If you are running the application from inside src/, install the dependencies from src/requirements.txt instead.

The application reads settings from a .env file through Pydantic settings. At minimum, provide:

APP_NAME=DRP Segmentation
APP_VERSION=1.0.0
GENERATION_BACKEND=cohere
COHERE_API_KEY=your_api_key_here
GENERATION_MODEL_ID=your_model_id_here

Optional tuning values:

INPUT_DAFAULT_MAX_CHARACTERS=2000
GENERATION_DAFAULT_MAX_TOKENS=2000
GENERATION_DAFAULT_TEMPERATURE=0.1
ORGINAL_LANGUGE=ar
DEFAULT_LANGUGE=ar

The Docker image starts the service with Uvicorn on port 7860.

uvicorn src.main:app --host 0.0.0.0 --port 7860

Upload a retinal image and receive:

• original_image
• overlay_image
• report
• full_prompt

The endpoint rejects non-image uploads.

Generates a structured report from sample model results using the configured LLM provider.

The notebooks in notebooks/ are still useful for experimentation and evaluation, but they are no longer the primary way to run the project.

• 01_disease_segmentation.ipynb - DR grading and preprocessing
• 02_vessel_segmentation.ipynb - vessel segmentation experiments
• 03_binary_classification_filter.ipynb - binary DR filtering
• 04_dr_grading_model.ipynb - grading model work
• 05_evaluation.ipynb - evaluation experiments
• 06_gradcam_pipeline.ipynb - explainability pipeline

• Keep raw datasets outside version control
• Store processed artifacts separately from source code
• Place the trained weights where src/main.py expects them before starting the API

• The codebase currently mixes API, prompting, and inference concerns, so the README now focuses on how to run and use the service.
• Several notebook workflows are still useful, but they are historical support material rather than the main entry point.