Main Projects

    In the following I classify my work in methods (meaning the new theoretical methods I personally developed) and applications. I describe some of my main projects in separate sections (for all my other projects see the publication list here). Even if in different fields, these projects are often connected and can benefit from each other.


Deformable Models

Segmentation into different, meaningful, areas in the image is almost always the most difficult part of image processing, and have to be specifically tailored to a narrow class of images. Snakes, or active contours, are deformable models used to find important divisions in an image, and genetic algorithms is an optimization method. Genetic snakes is a combination of the above subjects. Firstly I implemented this method during my Ph.D. studies, then I extended it in several directions (e.g. considering new energy terms for color images, and combinations of several snakes to handle complex contours). Genetic snakes has been successfully used to segment images in many fields. Among them, they have been used to segment the Foveal Avascular Zone in fundus ocular images obtained with Scanner Laser Ophthalmoscope, to segment liver in Computer Tomography scans of the abdomen, and more recently to segment bones in radiographic images of hands for bone age assessment.

Segmentation Algorithms

Clustering algorithms are useful in many applications, where the goal is to partition an image into homogeneous regions according to various criteria such as for example gray level, color, or texture. My main contribution to this field is a method that combines a fuzzy clustering algorithm, known as fuzzy c-means, with genetic algorithms. The objective function of the basic algorithm has been also modified to take into account the spatial information of image data and the intensity inhomogeneities. The genetic fuzzy c-means algorithm has been successfully applied to digital color camera photographs of beef meat, in order to classify different substances as muscle, fat and connective tissue; and to histological section images of bone implants, to segment bone and not bone structures.
I have been working on a new segmentation algorithm based on an evolutionary approach, which is reminiscent of the Life game. The evolution involves the colonization of a bi-dimensional world (i.e. the image) by a number of populations. Applications to microscopy, magnetic resonance and other images have been studied with very promising results.

Morphological Classification

In many medical image classification problems, the pathologist can not always explain why he reaches a particular diagnosis, but he takes his decisions based on experience acquired by observation of past cases. I designed a method that uses mathematical morphology to quantify aspects of the geometrical structure of images in a way that agrees with human intuition and perception. Genetic algorithms are used to optimize morphological operators. The novelty of the method is the use of morphological filters for image classification. The approach is based on performing the same operation sequence on images belonging to different classes, in order to find a filter (or a sequence of them) that leaves unaltered images of one class while changing the others. The whole image classification is divided in two step: a learning phase to extract parameters, and an implementation phase, in which the parameters are used to filter all the images and classify them. This method has been applied in the classification of microscopic images of breast tissue samples, to identifying benign from malignant changes.


Retina Images

This is the main topic of my Ph.D. studies. The goal of this project was to analyze retina images acquired with a Scanner Laser Ophthalmoscope, with focus on early signs of diabetic retinopathy, and in particular to Foveal Avascular Zone morphological changes.
My research included the development and the implementation of registration and enhancement techniques for the recovery of the vascular network from angiographic image sequences, and the realization of an automatic system to aid medical diagnosis, using segmentation, shape analysis and classification techniques. The segmentation is based on genetic snakes; the shape analysis uses features (region and boundary moments) which are tailored for the diagnostic purpose; the classification is based on a neural network-based classifier, optimized for the problem, which is able to perform feature selection at the same time as learning. The network architecture is also evolved and designed in a biologically plausible way.

SLO image of ocular fundus

Food Images

A project to develop sustainable food production methods has been the main topic of my Post-Doc at the Centre for Image Analysis, Uppsala. I have been working with camera photographs and Magnetic Resonance images of beef meat. Segmentation algorithms have been optimized for these kinds of images, in order to classify different substances as muscle, fat and connective tissue. The method for color camera photographs combines a fuzzy clustering algorithm and a genetic algorithm. The segmentation algorithm used for the Magnetic Resonance images includes also a filtering technique to remove intensity inhomogeneities, caused by non-uniformities of magnetic fields during acquisition. Moreover, I developed a method to measure homogeneity of fat distribution; this method is simple but accurate and gives a description of feature distribution and a measure of homogeneity, depending on both size and spatial organization of features, without requiring any individual measures of them. I studied a new method to predict fat content in meat images based on fractal theory. A web-based survey has been carried out to understand the choice of consumers when buying beef meat and compare preferences to features measured by image analysis.
The aim of the project on pig meat images was the quantification of pores, and the study of the relationship between the number and size distribution of pores and process condition and genotype. Image analysis in conjunction with fractal models has proved to be very suitable to describe pore distribution according to the sensory perception of the presence of pores.

digital camera picture of pork meat digital camera picture of beef meat NMR image of beef meat

Histological Images

I have been the leader of the medical image processing project for breast cancer detection, as a guest researcher at the Department of Technology, Örebro. The goal of the project was to develop a method to classify microscopic images of breast tissue samples and identifying benign from malignant changes. The method includes segmentation of epithelial cells and their nuclei, feature extraction and classification. The segmentation combines a fuzzy clustering algorithm with mathematical morphology. The extracted features range from size and shape measurements to fractal dimension. Features based on new morphological operators have being discovered by using genetic algorithms. Neural networks have been optimized for the specific problem and used in the classification.
The other project, were I was involved as leader, was on histological section images, for quantifications of bone tissue reactions to biomaterials. The objective was to measure percentages of contact between bone and inserted implants. I developed an automatic system for these measurements. The method includes image preprocessing (to deal with non uniformities due to the microscope illumination field), segmentation of bone and not bone structures, automatic selection of interesting areas (threads), and measurements of areas and perimeters.

normal breast tissue cancerous breast tissue bone implant

Forensic Images

The aim of my project at the European Centre for Soft Computing was to design an automatic procedure to aid the forensic anthropologist in the identification task of a missing person by putting into effect the photographic superimposition (which aims to overlay a model of the skull found and a photograph of the missing person). This procedure is based on the use of Soft Computing techniques and it is composed of three different stages: first a 3D model of the skull is reconstructed from different views acquired with a range scanner, second evolutionary algorithms and other metaheuristics are used for image superimposition, and third an automatic fuzzy decision system will be considered for the decision process. The first stage includes the use of the Scatter Search algorithm and of a local optimizer. I also developed an automatic feature extraction procedure for 3D range images of skulls. These features has been used to improve the registration of multiple views, and can be suitable for other applications as building anatomical atlas and identification of craniometrical landmarks. The second stage performs the superimposition of the 3D skull model and the 2D face photograph by finding the best fit between two sets of corresponding landmarks.