top of page

Reflectance confocal microscopy imaging


My research in medical imaging was conducted in close collaboration with the Pierre Fabre Dermo-cosmetic research center. The work carried out is mainly interested in the analysis and statistical modeling of reflectance confocal microscopy (RCM) images in order to characterize them for the identification of Lentigo. Reflectance confocal microscopy is an increasingly used medical imaging modality. Indeed, it allows to have a very good resolution of the skin (of the order of µm on the surface and in depth) without altering the skin. Its ease of use and speed of acquisition contribute to the strong development of this tool. However, a long learning curve is required before dermatologists can fully use the possibilities of this technique for diagnostic purposes. This is why more and more methods are used to minimize this learning time by automating some of the steps necessary for diagnosis. In this research, we want to learn how to characterize and distinguish healthy skin from skin affected by lentigo. Lentigos are age spots that appear mainly on the hand or on areas most often exposed to the sun. On the surface of the skin, they appear as a darker spot. Inside the skin, it is mainly at the dermis-epidermis junction that we will see differences. Indeed, the junction is larger and more disordered in an affected area. The motivation of my thesis was to propose statistical models to characterize RCM images to better identify lentigo affected skin from healthy skin. For this purpose, three main contributions have been proposed and described as follows:


  • A first contribution consisted in proposing a parametric statistical model to represent the texture of RCM images in the wavelet domain and then to classify these images into clusters using machine learning techniques. Specifically, it is a generalized Gaussian distribution whose scaling parameter is shown to be characteristic of the underlying tissue. [Halimi et al, Biomedical optics Express]


  • A second contribution consisted in proposing a statistical classification model adapted to the image domain for the characterization of tissues in RCM images using a new fast and robust estimation algorithm for the parameters of the generalized gamma distribution, based on a natural gradient approach. [Halimi et al, CAMSAP]


  • A third contribution was to propose a multiplicative noise observation model to explain the generalized gamma distribution of the data. Parametric Bayesian inference methods were then developed with this model to allow joint reconstruction and classification of skin reflectance confocal microscopy images. [Halimi et al, EUSIPCO]


The proposed models were tested on real RCM images collected during a study named cf010 conducted on 50 patient volunteers.

bottom of page