Enhanced Display of Mitoses in Hematoxylin-Eosin Digital Pathology Images

Abstract: Identifying mitotic figures in digital pathology images is useful in diagnosis and prognosis of cancer. The task itself is time consuming and sometimes difficult for pathologists. Recent research has focused on machine learning for automatic detection of mitoses. The aim of this work was to apply image processing to facilitate ocular examination of mitotic figures in hematoxylin and eosin stained images, since such an approach can be adopted faster in clinical applications. Several contrast adjustment algorithms were considered including contrast limited adaptive histogram equalization (CLAHE) and local Laplacian filters. The classic sharpening method unsharp masking was also used. Processing was done in the HSV and L'a'b' colour spaces and to channels of images where the stains hematoxylin and eosin had been separated through colour deconvolution. For validation a senior pathologist scored images with respect to the details of mitoses and general image quality. In the scoring unsharp masking obtained the best results with significant improvements of both the details of mitotic figures and general image quality. Local Laplacian filters and CLAHE improved the general image quality but only slightly improved the details of mitotic figures. CLAHE occasionally resulted in unwanted darkening of both mitoses and the image as a whole. The same algorithm applied to the hematoxylin channel of images showed potential although it requires careful estimation of the colour matrices that are used to deconvolve histopathology images into stain components. No investigated algorithm can alone eliminate the pathologists difficulties concerning mitotic figures. However, image enhancement may reduce the time needed for examination.