Training echo state networks for rotationinvarian

Authors: Philipp Kainz Harald Burgsteiner Martin Asslaber Helmut Ahammer

Publish Date: 2016/09/21

Volume: 28, Issue: 6, Pages: 1277-1292

PDF Link

Abstract

The main principle of diagnostic pathology is the reliable interpretation of individual cells in context of the tissue architecture Especially a confident examination of bone marrow specimen is dependent on a valid classification of myeloid cells In this work we propose a novel rotationinvariant learning scheme for multiclass echo state networks ESNs which achieves very high performance in automated bone marrow cell classification Based on representing static images as temporal sequence of rotations we show how ESNs robustly recognize cells of arbitrary rotations by taking advantage of their shortterm memory capacity The performance of our approach is compared to a classification random forest that learns rotationinvariance in a conventional way by exhaustively training on multiple rotations of individual samples The methods were evaluated on a human bone marrow image database consisting of granulopoietic and erythropoietic cells in different maturation stages Our ESN approach to cell classification does not rely on segmentation of cells or manual feature extraction and can therefore directly be applied to image dataThe initial step of diagnostic work in histopathology is the assessment of cellularity in context of tissue architecture Especially the diagnosis of bone marrow specimen requires a valid interpretation of different cell types with respect to their local distribution Cell types of hematopoiesis the process of blood stem cell maturation are categorized into granulopoiesis erythropoiesis and megakaryopoiesis which refer to maturation of white blood cells WBC red blood cells RBC and megakaryocytes respectively 2Samples of hematopoietic cell nuclei in the human bone marrow at 40times magnification and stained with Hematoxylin–Eosin HE Subsequent maturation stages of a granulopoietic cell a myelocyte b metamyelocyte and c band cell Especially in early stages where the cells are not fully differentiated different cell lineages share morphological characteristics d myelocytes and e orthochromatic normoblasts erythropoietic cellsAs a consequence both inter and intraobserver variability can be considerable affecting the accurate diagnosis of reactive or even premalignant and early malignant changes Thus automated image recognition systems exhibiting low variance high classification accuracy and predictable error are highly desirable for repeatable quantitative diagnostics 16 Since virtual microscopy using whole slide images scanned at high magnification is emerging to a standard in pathology departments 1 computeraided pathology using automated image analysis systems can easily be implemented in the routine diagnostic processOver the recent years a remarkable amount of research has been conducted on blood cell counting segmentation and classification in histopathological images for various applications Motivated by the aggressiveness of blood cancer and the requirement for early diagnosis most works were related to leukemia research in particular identifying different types of leukemia by classifying WBC from histopathology images of peripheral blood smear 5 17 27 36 39 40 42 43 44 48 52 or bone marrow obtained by aspiration 12 35 38 41 46 49 50 51 59 60 and trephine biopsy 3 Particularly some work focused on classification of WBC in healthy tissue 5 20 39 while others dealt with detecting pathological alterations from morphological characteristics of cells 12 41 43 59 60Notably a vast majority follows a conventional pattern recognition approach and used distinct steps for cell detection segmentation extraction of rotation and translation invariant features and classification Some studies mainly addressed detection and segmentation and relied on standard image processing techniques such as Hough space analysis 52 watershed transform Gabor filters and adaptive thresholding 20 or intensity clustering 59 60 Others pursued supervised learningbased cell detection approaches using feedforward neural networks FFNN 45 fuzzy cellular neural networks 44 and random forests RFs 26 Several approaches used statistical pattern recognition and classification techniques 22 such as support vector machines SVM 9 17 35 46 FFNN 17 19 27 31 36 49 50 51 and Bayesian classifiers 6 42 43 to learn feature vectors representing individual cell objects Decision treebased methods such as regression trees 3 hierarchical trees using genetic algorithms for node optimization 56 or RFs 7 12 41 were used as well as knearest neighbor 37 38 41 43 or heterogeneous classifier ensembles 12 37 Employing hierarchical models has been shown to be more powerful than using singlestage classifiers 46 48Features related to shape and texture of cell nuclei were most frequently used and seemed to provide more discriminative power than statistical features from intensity histograms This is reasonable since—despite proper histopathological staining protocols—nuclei of different cell classes share very similar intensity patterns after staining 40 cf Fig 1 However the choice and importance of features depend on the application For instance it was shown that features computed from cytoplasm can even be omitted for WBC classification and that the problem can be downscaled to using features from cell nuclei only 50 In the context of another application using features from both nuclei and cytoplasm resulted in higher classification performance 13 Recent work of Reta et al 41 on bone marrow cells concluded that features extracted from nuclei and cytoplasm separately are more discriminative than features from entire cells In previous studies the total number of features varied from a small set of four to over 190 comprising objectlevel features as well as global image features such as wavelet coefficients 36 Nevertheless handcrafting features from images require prior knowledge and experience they are not easily transferable to other problems and may as well remove significant information or introduce nondiscriminative information Thus authors of previous papers frequently extracted feature candidates and applied automatic feature selection procedures to extract the most significant subset and hence compress the available information to achieve a better generalization performance 17 37 39 42 It has been shown that this strategy generally improved the classification results compared to using all available features 35 42 46 on specific problems On the other hand working directly on image intensity data provides a directly observable object representation that is not influenced by errors of preceding segmentation steps that are frequently inevitably to extract objectlevel features Nevertheless only a minority of previous work focused on learning a classifier from raw cell images 17 44 58 but reported promising resultsVery little work has been reported on quantitative analysis of bone marrow trephine biopsy images 3 or quantification of blood cell maturation 20 Tissue microarchitecture is usually well preserved after histological preparation in bone marrow trephine biopsy samples At the proper magnification and using suitable histological staining protocols this enables to inspect the morphological differences among subsequent maturation stages but also introduces and emphasizes background structures irrelevant to cell classification The most common stains used for tissue specimen were May–Grünwald–Giemsa MGG HE bone marrow and Wright peripheral blood since morphological characteristics of objects of interest can be well represented Featurebased discrimination of cells is usually less complicated in smear images depicting differentiated cells than in trephine biopsies segmentation methods can more easily be applied to the cell objects without getting distracted by heterogeneous background Despite the efforts of previous work several issues have not yet been addressed and the quantification of blood cell maturation in the bone marrow has not been sufficiently studied yet

Keywords: