Cell segmentation and classification via unsupervised shape ranking

As histology patterns vary depending on different tissue types, it is typically necessary to adapt and optimize segmentation algorithms to these tissue type-specific applications. Here we present an unsupervised method that utilizes cell shape cues to achieve this task-specific optimization by introducing a shape ranking function. The proposed algorithm is part of our Layers™ toolkit for image and data analysis for multiplexed immunohistopathology images. To the best of our knowledge, this is the first time that this type of methodology is proposed for segmentation and ranking in cell tissue samples. Our new cell ranking scheme takes into account both shape and scale information and provides information about the quality of the segmentation. First, we introduce cell-shape descriptor that can effectively discriminate the cell-type´s morphology. Secondly, we formulate a hierarchical-segmentation as a dynamic optimization problem, where cells are subdivided if they improve a segmentation quality criteria. Third, we propose a numerically efficient algorithm to solve this dynamic optimization problem. Our approach is generic, since we don´t assume any particular cell morphology and can be applied to different segmentation problems. We show results in segmenting and ranking thousands of cells from multiplexing images and we compare our method with well established segmentation techniques, obtaining very encouraging results.