An improved matched filter for blood vessel detection of digital retinal images

An improved matched filter for blood vessel detection of digital

retinal images

Source: Computers in Biology and Medicine (2007) pp. 262 – 267

Authors: Mohammed Al-Rawi, Munib Qutaishat, and Mohammed Arrar

Impact factor: 1.068 (2006)

Reporter: Kai Hung Chen

Date: Mar 11, 2008

Outline

Introduction

Proposed method

Experiment results

Introduction

(a) : The green band of a digital retina image

(b) : The profile of a one pixel width at the 200th row of the retina image shown in (a)

Proposed Method

2 2

exp, 2

2Lyxyxk

L: the length of the vessel segment that has the same orientation

σ: the spread of the intensity profile

Proposed Method

12 ..., ,2 ,1 wherecossin

sincos

iyxvupi

The point Pi that belongs to N is:

3 where, 2 , ,, TLuTuvuN

Neighborhood N is defined as:

Proposed Method

Npuyxk ii 2

exp, 2

2

The corresponding weights in the kernel i ( i =1, . . . , 12 which is the number of kernels) are given by:

.in points ofnumber theis

,,1 where,,,'

Na

yxkammyxkyxk iNpiiii i

The filter is normalized as:

Quality Factor

True pixels: pixels detected as vessels and they appear as vessels in the hand label image.

False pixels: pixels detected as vessels yet they appear as non-vessels in the hand labeled image.

true_ratio: divide the true pixels by the number of vessel pixels in the hand labeled image.

false_ratio: divide the false pixels by the number of non-vessel pixels in the hand labeled image.

Quality Factor ( QLσT ) = true_ratio − false_ratio

Experiment Results

Experiment Results

Experiment Results

MAA: maximum accuracy average, the average accuracy of all images.

Experiment Results

Experiment Results

Experiment Results

Experiment Results