Retinal Vessel Segmentation — CLAHE + Matched Filter

Segmentasi pembuluh darah retina menggunakan pipeline image processing (CLAHE + Matched Filter) dan klasifikasi anomali retina menggunakan Machine Learning (SVM / Random Forest).

Pipeline

RGB Image → Green Channel → FOV Mask → Illumination Correction
         → CLAHE → Gaussian → Matched Filter (12 orientations)
         → Otsu Thresholding → Morphological Post-processing
         → Binary Vessel Mask

Klasifikasi Anomali (ML):

Vessel Mask → Ekstrak Fitur Vessel → SVM / Random Forest → Normal / Abnormal (DR)

Fitur:
  • Vessel Density       — rasio pixel vessel terhadap FOV
  • Mean Vessel Width    — rata-rata lebar vessel via skeletonization
  • Tortuosity           — rasio panjang lengkung / jarak lurus per segmen
  • Branching Points     — kepadatan titik percabangan vessel
  • Fractal Dimension    — kompleksitas pola vessel (box-counting)

Struktur Proyek

src/retina_seg/
├── dataset/        — loader untuk CHASE_DB1 dan APTOS 2019
├── preprocessing/  — green channel, FOV mask, illumination correction
├── enhancement/    — CLAHE, Gaussian, Matched Filter
├── segmentation/   — Otsu thresholding, morphological post-processing
├── evaluation/     — metrik SE/SP/ACC/AUC, overlay visualisasi
└── ml/             — ekstrak fitur vessel, train/predict SVM & RF
app.py              — Streamlit demo (3 tab)
tests/              — 28 unit tests

Installation

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Dataset Setup

Segmentasi vessel (batch evaluation):

• CHASE_DB1 — lihat data/CHASE_DB1/README.md

Klasifikasi anomali (ML training):

• APTOS 2019 Blindness Detection — download dari: https://www.kaggle.com/competitions/aptos2019-blindness-detection/data
• Ekstrak ke data/APTOS2019/ dengan struktur:

  data/APTOS2019/
  ├── train_images/   ← ~3.662 gambar .png
  └── train.csv       ← kolom: id_code, diagnosis (0–4)
  

Demo App

streamlit run app.py

Tab	Fungsi
Single Image	Upload gambar → lihat tiap stage pipeline → download segmentasi
Anomaly Classification (ML)	Train SVM/RF pakai APTOS 2019 → prediksi Normal/Abnormal
Batch Evaluation	Evaluasi CLAHE+MF pada seluruh CHASE_DB1 → tabel SE/SP/ACC/AUC

Python API

import cv2
from retina_seg.preprocessing.pipeline import run as preprocess
from retina_seg.enhancement.pipeline import run as enhance
from retina_seg.segmentation.pipeline import run as segment
from retina_seg.ml.pipeline import extract_image_features, train, predict

img_rgb = cv2.cvtColor(cv2.imread("fundus.jpg"), cv2.COLOR_BGR2RGB)

# Segmentasi vessel
corrected, fov_mask = preprocess(img_rgb)
enhanced = enhance(corrected)
vessel_mask = segment(enhanced, fov_mask)

# Klasifikasi anomali (setelah model ditraining)
features = extract_image_features(vessel_mask, fov_mask)  # shape (5,)
# model, scaler = train(X, y, method="random_forest")
# result = predict(model, scaler, vessel_mask, fov_mask)
# → {"label": 1, "label_str": "Abnormal (DR)", "probability": 0.87, ...}

Run Tests

pytest tests/ -v

Hasil Evaluasi — CHASE_DB1 (28 gambar)

Metrik	Mean	Std
SE (Sensitivity)	0.7687	±0.0542
SP (Specificity)	0.8463	±0.0490
ACC	0.8380	±0.0399
AUC	0.8560	±0.0191