Texture Classification Using Neural Networks and Local Granulometries

This paper presents a method for segmenting interstitium and tubules in images of kidneys' biopsies. Openings by structuring elements of increasing size, forming a granulometry, were performed on the entire image. For every pixel x and for each size of the structuring element the volume over a small window centered at x was measured (a local Granulometry). The vectors deened as the volume gradient served as an entry to a neural network (NN). The NN was taught to discriminate between vectors corresponding to pixels of the interstitium (textured region) and vectors correspondingto pixels of the tubules (non-texturedregion). The correlationfactor between the area of the interstitium and the renal function was computed and compared to the results obtained with the manual procedure and two other automatic procedures.