Segmentation of blood vessels in retinal fundus images Healthy - PowerPoint PPT Presentation

Segmentation of blood vessels in retinal fundus images

Healthy Hypertension damage

Ophtalmoscopy

Retinal image Segmentation Automatic segmentation

Simple bar-selective fjlter: B- COSFIRE 𝜍 Automatic confjguration f Each point described by: Rotation invariance:

Filter application • Use a Gaussian for tolerance, std. Dev.: • Response for one point: • Multiply the shifted responses -> COSFIRE

Pre-processing Original image Mask Green channel

Putting it all together B-COSFIRE Threshold

T uning parameters B-COSFIRE • σ: • ρ: The largest circle 𝜍 • σ 0 • α Symmetric: σ = 4.8, ρ = 20, σ 0 = 3, α = 0.3 Assymetric: σ = 4.4, ρ = 36, σ 0 = 1, α = 0.1

Segmentation performance t = [0,1] t TPR,FPR

IOSTAR EasyScan Optics B.V. The Netherlands

Results

Sensitivity of parameters Symmetric: σ 0 = 3 σ = 4.8 α = 0.3 Paired T-test ed: Signifjcantly difgerent segmentation White: Similar segmentation Small Large performance difgerence deviation

Machine learning approaches Training pase Working phase

Deep neural network B-COSFIRE Training/estimating > ± 10 8 hours minutes Single GPU Human Segmenting > 10 seconds 92 seconds 2 GHz CPU High-end GPU > AUC: .9720 AUC: .9614

Thank you.