DEEP VESSEL TRACKING: A GENERALIZED PROBABILISTIC APPROACH VIA DEEP LEARNING

Automatically segments 2D retinal vessels using N4 fields, CNN, and probabilistic vessel tracking

Automatically segments 2D retinal vessels by:
-Sampling retinal image windows with their corresponding manual segmentations
-Training the image windows to a PCA-reduced vector of the manual segmentations using convolutional neural networks (CNN)
-Building a dictionary of by transforming a new set of sampled windows into the PCA-reduced vectors using the (CNN) and tying them to their corresponding manual segmentations
-Segment a new retinal image window pixel-by-pixel by:
Passing all the windows that include that pixel into the trained network
Searching the dictionary for the nearest neighbor of the resulting vectors
Acquiring the corresponding manual segmentations
Taking the pixel in the manual segmentations that corresponds to the location of the target pixel
Averaging the acquired pixel values
-Segment the entire retinal image window-by-window using probabilistic tracking methods that search the already-segmented-areas to select the next segmentation window
Improves previous vessel segmentation methodology:
-Improved accuracy in vessel segmentation using N4 Fields compared to other methods
-Improved speed in vessel segmentation using vessel tracking compared to a full segmentation using N4 Fields. This is because not every pixel is segmented, and a single pixel takes a considerable amount of time to segment.
Contact me at [e-mail] if you want result examples or have questions (the code can be confusing)