|
Colposcopy is a diagnostic method used to detect cancer precursors and cancer of the uterine cervix. Computer-Aided-Diagnosis (CAD) for colposcopy is a new field in medical image processing. Colposcopists analyze glare (glint or specular reflection) patterns on the cervix to assess the surface contour (3D topology) of lesions, an important feature used to evaluate lesion severity. However, glare in the imagery presents major problems for automated image analysis systems. Glare eliminates all information in affected pixels and can introduce artifacts in feature extraction algorithms, such as acetowhite region detection. Although cross-polarization filters can be used to eliminate glare, the reality is that we have to deal with glare when we want to use existing cervical image databases or use an instrument that does not provide cross-polarized imagery. Here, we present the details and preliminary results of a glare removal algorithm for RGB color images of the cervix that can be used as a pre-processing step in CAD systems. The algorithm can be extended to multispectral and hyperspectral imagery. The basic approach of the algorithm is to extract a feature image from the RGB image that provides a good glare to background ratio, to detect the glare regions in the feature image, to extend the glare regions to cover all pixels that have been affected by the glare, and to remove the glare in the affected regions by filling in an estimate of the underlying image features. In our current implementation we use the green (G) image component as the feature image, given its high glare to background ratio and simplicity of calculation. Glare regions are either detected as saturated regions or small high contrasted bright regions. Saturated regions are detected using an adaptive thresholding method. Small high contrasted bright regions are detected using morphological top hat filters with different sizes and thresholds. The full extent of the glare regions is estimated by using a morphological constraint watershed segmentation to find the contour of the glare regions and adding a constant dilatation. The image features are estimated by interpolating the R,G,B color components individually from the surrounding regions based on Laplace’s equation and modifying the intensity (I) component of the HSI color space transformed image. As the glare pattern is important to the physician, we embed it as a just visible intensity texture that does not affect the image processing. The performance of the algorithm is demonstrated using human subject data.
|
