Contribution to the development of multispectral image analysis strategies for automatic and rapid diagnosis of malaria ; Contribution au développement des techniques d’analyse d’images multispectrales pour l’aide au diagnostic du paludisme

Malaria is an infectious and potentially fatal disease which affects nearly half of the world’s population. It kills one person every 30 seconds in endemic area. An effective surveillance through early and accurate diagnosis is essential to fight against this pandemic.This work proposes a new malaria diagnosis methodology based on multispectral imaging.Our first contribution is an algorithm that reduces the acquisition time of blood cell images while maintaining the normalization preprocessing. The normalization is an essential pre-processing step in the diagnostic which ensure uniformity of brightness in the images. The method is based on estimation of the Bright reference image which represents luminosity and contrast variability function from the background part of the image. This is accomplished by two different re-sampling methodologies namely gaussian random field simulation by a geostatistical method (variogram analysis) and Bootstrap re-sampling. To avoid intensity saturation problem of some pixels, we addressed the electronic noise by Hampel’s Outlier Detection and Imputation Method. The proposed solution is very fast and takes less than 20 seconds.The second contribution is about an erythrocyte detection and segmentation process; taking advantage of Beer Lambert's law by using first statistical standardization equation applied to transmittance, the local adaptive threshold algorithms, watershed algorithm and contour closure by hysteresis.To identify parasitized RBCs, a classification process is performed on 12 textures descriptors of segmented cells for 13 wavelengths and three geometries (Scattering, Refection and Transmission). A multivariate functional principal component analysis was performed on the data and then unsupervised classification, by K-means partitioning algorithms and hierarchical ascending classification (CAH) allowing to isolate the classes of parasitized and healthy cells. Finally, the supervised classification is done by machine learning algorithms.To summarize, our process reduced ...