Modelisation of the noise and detection of blood vessels in fundus retina images

• Autores: Mariem Ben Abdallah
• Directores de la Tesis: Julio Esclarín Monreal (dir. tes.)
• Lectura: En la Universidad de Las Palmas de Gran Canaria ( España ) en 2016
• Idioma: español
• Tribunal Calificador de la Tesis: Rached Tourki (presid.) , Agustín Trujillo Pino (secret.) , Ahmed Ben Hmida (voc.) , Máximo Méndez Babey (voc.)
• Texto completo no disponible (Saber más ...)
• Resumen

  • For decades, retinal images are widely used by ophthalmologists for the detection and follow-up of several pathological states. These images directly provide information about the normal and abnormal features in the retina. The normal features include the optic disk, fovea and vascular network. There are different kinds of abnormal features caused by Diabetic Retinopathy (DR) such as microaneurysm, hard exudate, soft exudate, hemorrhage and neovascularization.

    Automated analysis and interpretation of fundus images has become a necessary and important diagnostic procedure in ophthalmology. The extraction of blood vessels from retinal images is an important and challenging task in medical analysis and diagnosis. In this work a Modelisation of the Noise and Detection of Blood Vessels in Fundus Retina Images is presented.

    The use of the rough images is not desirable if one wishes to detect automatically any information from the retinal image because it contains some noise. Thus, an image pre-processing is required before any treatment. In image pre-processing by partial differential equations, the first and simplest models to have and use are based on linear diffusion. The common difficulty of linear filters is the excessive smoothing that makes track edges difficult. Therefore, we can affirm that any improvement of these linear models must be carried out inside the diffusion operator, thus sacrificing their linearity. The work achieved in this context will make the subject of the following chapter 1. This part of thesis treats the automatic preprocessing of retinal vascular network in fundus images, using various versions of anisotropic diffusion filters, in order to improve their interpretation. To evaluate the chosen methods, we have performed image enhancement parameters, mean preservation and variance reduction, and edge preservation.

    To reduce image noise, the SRAD, a version of anisotropic diffusion scheme used the noise model with some assumptions such as to be additive, uniform and independent of the RGB image data. Afterwards, this approache cannot effectively recover the ”true” signal (or its best approximation) from these noisy acquired observations. Since, noise estimation from a single retinal image is required in order to improve the denoising filter. To compute the Noise Level Function (NLF), an iterative noise estimation process is presented in the chapter2. This method smooth out the noisy image by convolving it with a low pass filter, and then partition the smoothed image into homogeneous regions with both similar spatial coordinates and RGB pixel values. Then, it estimates the noise-free signal and the noise variance for each region and forms the scatter-plots of samples of noise variances on the estimated noise free signals of each RGB channel. The lower envelope strictly and tightly below the sample points is the estimated NLF curve. However, the estimated variance of each region is an over-estimate of the noise level because it may contain the signal, so the obtained lower envelope is not the true bound estimate of the NLF. To solve that, an inference problem in a probabilistic framework was formulated to derive the real bound on the image noise level.

    The chapter 3, presents a new Adaptive Noise Reducing Anisotropic Diffusion (ANRAD) method to improve the image quality, which can be considered as a modified version of a Speckle- Reducing Anisotropic Diffusion (SRAD) filter. The SRAD works very well for monochrome images with speckle noise. However, in the case of images corrupted with other types of noise, it cannot provide optimal image quality due to the inaccurate noise model. The ANRAD method introduces an automatic RGB noise model estimator in a partial differential equation system similar to the SRAD diffusion, which estimates at each iteration an upper bound of the real noise level function by fitting a lower envelope to the standard deviations of per-segment image variances. Compared to the conventional SRAD filter, the proposed filter has the advantage of being adapted to the color noise produced by today's CCD digital camera. The simulation results show that the ANRAD filter can reduce the noise while preserving image edges and fine details very well. Also it is favorably compared to the fast non-local means filter, showing an improvement in the quality of the restored image. A quantitative compare is on measure is given by the parameters like the mean structural-similarity index and the peak signal-to-noise ratio.

    In the chapter 4, an implementation of our new anisotropic diffusion filter is applied in order to restore connected vessel lines and eliminate noisy lines. Next, a multi-scale line-tracking procedure allows detecting all vessels having similar dimensions at a chosen scale. Computing the individual image maps for one scale requires different steps. First, a number of points are pre-selected using the eigenvalues of the Hessian matrix. These points are expected to be near to a vessel axis. Then, for each pre-selected point, the response map is computed from gradient information of the image at the current scale. Finally, the multi-scale image map is derived after combining the individual image maps at different scales (sizes). Two publicly available datasets have been used to test the performance of the suggested method. The main dataset is the STARE Project’s dataset and the second one is the DRIVE dataset. The experimental results, applied on the STARE dataset, show a maximum accuracy average of around 93:88%. Also, when performed on the DRIVE database, the maximum accuracy average reaches 93:89%.