Stain separation, cell classification and histochemical score in digital histopathology images

• Main Author: Liu, Jingxin
• Published: University of Nottingham 2018
• Subjects: T Technology (General)
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.757543

This thesis focuses on developing new automatic techniques addressing three typical problems in digital histopathology image analysis, histochemical stain separation at pixel-level, cell classifications at region level, and histochemical score assessment at image level, with the aim of providing useful tools to help histopathologists in their decision making. First, we study a pixel-level problem, separating positive chemical stains. To realise the full potential of digital pathology, accurate and robust computer techniques for automatically detecting biomarkers play an important role. Traditional methods transform the colour histopathology images into a gray scale image and apply a single threshold to separate positively stained tissues from the background. In this thesis, we show that the colour distribution of the positive immunohistochemical stains varies with the level of luminance and that a single threshold will be impossible to separate positively stained tissues from other tissues, regardless how the colour pixels are transformed. Based on this observation, two novel luminance adaptive biomarker detection methods are proposed. The first, termed Luminance Adaptive Multi-Thresholding (LAMT) first separate the pixels according to their luminance levels and for each luminance level a separate threshold is found for detecting the positive stains. The second, termed Luminance Adaptive Random Forest (LARF) applies one of the most powerful machine learning models, random forest, as a base classifier to build an ensemble classifier for biomarker detection. Second, we study a cell-level problem, the cell classification task in pathology images. Two different classification models are proposed. The first model for HEp-2 cell pattern classification comes with a novel object-graph based feature, which decompose the binary image into primitive objects and represent them with a set of morphological feature. Work on cell classification is further extended using deep learning model termed Deep Autoencoding-Classification Network (DACN). The DACN model consists of an autoencoder and a conventional classification convolutional neural network (CNN) with the two sharing the same encoding pipeline. The DACN model is jointly optimized for the classification error and the image reconstruction error based on a multi-task learning procedure. We will present experiment results to show that the proposed DACN outperforms all known state-of-the-art on two public indirect immunofluorescence stained HEp-2 cell datasets and H\&E stained colorectal adenocarcinomas cell dataset. Third, we study an image-level problem, assessing the histochemical score of a histopathology image. To determine the molecular class of the tumour, pathologists will have to manually mark the nuclei activity biomarkers by assigning a histochemical score (H-Score) to each TMA core with a semi-quantitative assessment method. Manually marking positively stained nuclei is a time consuming, imprecise and subjective process which will lead to inter-observer and intra-observer discrepancies. In this thesis, we present an end-to-end deep learning system which directly predicts the H-Score automatically. Our system imitates the pathologists' decision process and uses one fully convolutional network (FCN) to extract all nuclei region, a second FCN to extract tumour nuclei region, and a multi-column convolutional neural network which takes the outputs of the first two FCNs and the stain intensity description image as input and acts as the decision making mechanism to directly output the H-Score of the input TMA image. To the best of our knowledge, this is the first end-to-end system that takes a TMA image as input and directly outputs a clinical score. We will present experimental results which demonstrate that the H-Scores predicted by our model have very high and statistically significant correlation with experienced pathologists' scores and that the H-Score discrepancy between our algorithm and the pathologists is on par with the inter-subject discrepancy between the pathologists.