Epithelium and stroma segmentation using multiscale superpixel clustering.

Introduction: Accurate image segmentation is essential in quantitative histopathology although challenging due to tissue complexity, heterogeneity and the uncertainty of scene contents. Because structures detected at certain resolutions do not always coincide with the resolution at which other features are detected, we investigated a multiscale approach of image segmentation. Method: We developed an unsupervised framework using multiscale superpixels and k-means clustering. Images were partitioned into increasingly sized superpixels and clustered into 3 classes based on tissue stain uptake after colour deconvolution (aiming to identify background, epithelium and stroma) in H&E oropharyngeal cancer TMA sections. To overcome arbitrary label assignments during clustering, labelled regions were sequentially matched across consecutive scale image pairs using the additive inverse of the Dice index to generate a ‘cost matrix’ for all label combinations. A bipartite graph matching algorithm identified the combination that minimised the matrix overall cost. Results: From the matched ensemble, labelling probability images were generated to compute a ‘most likely’ segmentation. Fifty-six micro-array images compared with annotated standards achieved an average Dice index of 0.76 (median 0.79). The ‘most likely’ segmentations versus the standard, achieved a higher Dice index than those of the smallest and largest scale results (71% and 95% of instances respectively) and the average Dice over all scales (in 93% of instances). Conclusions: Our unsupervised segmentation, on average, performs better than single resolution superpixel clusterings. The method can also be applied to samples stained by other methods, once the dye RGB vectors (for colour deconvolution) have been appropriately determined.