Illumination field inhomogeneity strongly affects the visual appearance of an image. It has a major influence on automatic information extraction within an image and its correction is critical for comparison or model learning across images. In this work a unique medical repository of cervicographic images ("cervigrams") collected by the National Center Institute (NCI), National Institute of Health (NIH) is being addressed. The large diversity of cervix shapes within this database, as well as the acquisition set-up, lead to varying illumination conditions among and within the cervigrams, which hamper their automatic analysis. Illumination correction is therefore one of the first preprocessing steps required prior to the image analysis task. This paper presents a method for illumination correction in cervigrams based on a generalized expectation maximization (GEM) algorithm that interleaves pixels classification with estimation of class distribution and illumination field parameters. For cross-image analysis a normalization of the image dynamic range is conducted, using prior knowledge on cervix tissue intensity distribution. Experimental results are provided and evaluated on a set of 110 cervigrams that were manually labeled by an NCI expert. Unsupervised segmentation as well as initial supervised tissue classification results are presented.
展开▼
