 |
Giancardo Luca |
   |
| Project:
|
| |
| Supervisor:
|
| |
| Host University:
|
| |
| Personal introduction |
top |
Hello!
I am from Savona, an Italian town on the Mediterranean sea surrounded by the Alps. Since I was a 15 I have been programming algorithms for local companies and at the age of 18 I moved to Southampton (England) for a BSc in Software Engineering. Thanks to the subject of my undergraduate thesis I discovered the field computer vision which allowed me to be hired as a researcher by Real Time Tracking, a company that creates various vision based systems particularly for sport applications. After one year in there, I decided that I needed to know much more in order to create real cutting edge systems, therefore I took my trusted saxophone and I started the VIBOT adventure.
I cannot be more glad about the decision I took two years ago: I learned an incredible amount of things, I met wonderful people from many cultures, I had the chance of learning new languages and I worked for a National Laboratory in USA. What can I ask more?
|
|
| |
The World Health Organisation estimates that 135 million people have diabetes mellitus worldwide and that the number of people with diabetes will increase to 300 million by the year 2025. A great effort of the research community is geared towards the creation of an automatic screening system able to promptly detect diabetic retinopathy with the use of fundus cameras.
The key for low cost widespread screening is a system usable by operators with little training. For this reason an algorithm able to automatically judge the quality of the fundus image captured is a necessary preprocessing step for a correct diagnosis. Some methods that address this problem can be found in the literature but they have some drawbacks like accuracy or speed.
This thesis aims to develop and test a new method for automatically judging the image quality with a score from 0 to 1. To do so various descriptors are used, among them a new set of features independent of field of view or resolution to indirectly describe the morphology of the patient's vessels. A C++ implementation of the complete algorithm is developed and tests suggest that the image quality can be estimated in around 2 seconds, a time one order of magnitude shorter than previous techniques with the same or better accuracy.
|
|
| |
The quality assessment is approached as a pattern recognition / machine learning problem. A set of features independent of camera and resolution is extracted and used to train a support vector machine classifier:
Mask / Virtual Field of View
the mask is obtained by employing a region growing algorithm with four seeds at the edges of the fundus image (figures (a) and (b)). The mask edges are extracted with simple morphological operations obtaining figure (c). Then, the VFOV is calculated by fitting an ellipse to the mask edges (figure (d)). The VFOV allows to approximate part of the information lost by the fundus camera as shown in the left figure.
Vessel Segmentation
The green channel of the RGB fundus image is extracted and a standard vessel segmentation algorithm is used (proposed by Zana and Klein in 2001). The figure above shows various stages of the algorithm.
Elliptical Local Vessel Density
A series of windows built with an the elliptical reference of the VFOV are used to extract the local densities of the vessels. Each window is employed as a dimention of the feature vector. The following figure shows various examples of ELVD.
Saturation Histogram
The original image is converted to HSV space. 2 bins of the Saturation channel histogram are added to the feature vector and normalised by the mask area.
Classification
Datasets representative of the variation of fundus images in terms of quality, camera used and patients ethnicity are used to test the system. The training was done on a linear SVM with a leave-one-out validation.
|
|
| |
A perfect classification is obtained for all the images (around 100). The average score difference between good and poor is more than 0.9.
The time required to compute the quality of an image is ~2 seconds on a single core PIV machine.
Given the excellent results, a paper with this technique was accepted for the 30th Annual International IEEE EMBS Conference in Vancouver. The draft paper can be accessed in the following section.
|
|
| |
|
MSc Thesis Project documents:
|
|
