Skip to content

Latest commit

 

History

History
45 lines (34 loc) · 1.74 KB

README.md

File metadata and controls

45 lines (34 loc) · 1.74 KB
Raw

Blood vessel detection in eye fundus

Development

Please try to follow PEP8 style guide for Python. You can use autopep8 formatter and flake8 as linter for this purpose. Don't set any additional arguments for these tools. Also note that each vessel detection method has own directory.

Method 1 - Image Processing

Directory image_processing contains python script that uses image processing approach. Basically we can distinguish three stages of processing.

Pre-processing

  • conversion to grayscale (green channel)
  • histogram equalizaiton (CLAHE)
  • denoising (non-local means)

Vessel Detection

  • frangi ridge operator

Post-processing

  • thresholding
  • removing small elements (connected components with stats)
  • removing border (based on black color in hsv)

Method 2 - Machine Learning

Directory machine_learning contains jupyter notebook that uses machine learning approach. Basically we can distinguish five stages of processing.

Pre-processing

  • image resizing

Feature extraction

  • dividing image into small windows (view_as_windows)
  • calculating RGB mean, RGB standard deviation and image moments for each window - x values
  • deciding whether center of window contains vessel (255) or not (0) - y values

Learning

  • undersampling of background class
  • splitting data (train_test_split)
  • using random forest
  • tuning hyper-parameters (RandomizedSearchCV)
  • evaluating model accuracy

Classification

  • using prepared model to detect vessels in windows

Post-processing

  • building image from windows and predicted labels
  • removing small elements (connected components with stats) to improve quality