1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Title:	A combinatorial segmentation method of b-cosfire and frangi filters with random forest for retinal blood vessel detection
Authors:	Aziah Ali (P84141)
Supervisor:	Aini Hussain, Prof. Dr.
Keywords:	Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Diabetic retinopathy Fundus image Retinal diagnosis
Issue Date:	4-Jul-2021
Description:	Diabetic Retinopathy (DR) is one of the major eye diseases which may cause permanent blindness if not treated early. Fundus images have been widely used for routine retinal screening in most hospitals to detect DR symptoms in a timely manner, thus reducing the risk of total blindness. A computer-assisted diagnosis system (CADS) can be of significant assistance to ophthalmologists in performing retinal diagnosis. Automatic retinal blood vessel (RBV) segmentation is one of the important pre-requisites for a CADS to facilitate subsequent processes for accurate and efficient retinal diagnosis. Large RBVs are relatively easier to segment due to their high contrast against the retinal background. On the other hand, smaller RBVs are more challenging because of their narrow size and low contrast. Existing RBV segmentation methods perform well in detecting large RBVs but leave smaller RBVs mostly undetected. As such, an efficient segmentation method for detection of both large and small RBVs from fundus image is required. This is important for a CADS to enable computation of important RBV parameters such as length, diameter and tortuosity for diagnosis purposes. Thus, this research focuses on the main aim to develop an unsupervised combinatorial segmentation method for RBV detection based on the combination of B-COSFIRE (BC), Frangi (FR) and Background Normalisation (BN) filters with Random Forest (RF) classifier. This study involves data from three public fundus image databases, namely DRIVE (20 sets), STARE (20 sets) and HRF (45 sets) together with one local database HUKM (30 sets). The proposed method involves three main steps, namely preprocessing (PreP), RBV segmentation and post-processing (PoP). Colour fundus image is used as the input to PreP step where the green channel image (GCI) is extracted as a grayscale image (GSI) and pre-processed to highlight RBV structures. For segmentation, BC filter is first applied to the pre-processed GCI to detect large RBV producing a GSI, called GBC. Next, two filters, namely BN and FR, are used to enhance the GBC to highlight the small vessels and doing so, produces the GBN and GFR, respectively. Two hybrid features are considered in this study to fuse the large and small RBV features, namely GBC with GBN and GBC with GFR using two fusion methods namely Grayscale (Technique A) and Binary fusion (Technique B) to produce ABCBN, ABCFR, BBCBN & BBCFR. The PoP steps are then applied to the hybrid features in order to remove noisy pixels. A combinatorial method is proposed by training RF models using the fused hybrid feature to detect more vessel pixels (ABCBN+RF, ABCFR+RF ,BBCBN+RF & BBCFR+RF). The best performance based on mean Sensitivity (Sn) and mean Specificity (Sp) are obtained from the combinatorial method of BBCFR+RF (85.03%, 95.17%), followed by BBCBN+RF (84.03%, 96.07%),), BBCFR (81.29%, 96.13%) and BBCBN (80.51%, 96.09%). Its quantitative results indicate comparable performance to the stateof- the-art methods using Deep Learning (DL) and qualitatively, able to better detect small RBVs. Validation on the local dataset HUKM also shows the best qualitative results is obtained using BBCFR+RF. A Graphical User Interface is also developed for simple application of the proposed method on actual fundus images. In conclusion, this study has confirmed the effectiveness of BBCFR+RF combinatorial method to segment RBV from fundus image effectively. Thus, ensuring efficient computation of important RBV parameters of the CADS towards a faster and accurate diagnosis system for eye diseases.,Ph.D.
Pages:	184
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.