Computational Histopathology Analysis based on Deep Learning
View/ Open
Metadata
Show full item recordAbstract
Pathology has benefited from the rapid progress in technology of digital scanning during the last decade. Nowadays, slide scanners are able to produce super-resolution whole slide images (WSI), also called digital slides, which can be explored by image viewers as an alternative to the use of conventional microscope. The use of WSI together with the other microscopic and molecular pathology images brings the development of digital pathology, which further enables to perform digital diagnostics. Moreover, the availability of WSI makes it possible to apply image processing and recognition techniques to support digital diagnostics, opening new revenues of computational pathology. However, there still remain many challenging tasks towards computational pathology such as automated cancer categorisation, tumour area segmentation, and cell-level instance detection. In this study, we explore problems related to the above tasks in histology images. Cancer categorisation can be addressed as a histopathological image classification problem. Multiple aspects such as variations caused by magnification factors and class imbalance make it a challenging task where conventional methods cannot obtain satisfactory performance in many cases. We propose to learn similarity-based embeddings for magnification-independent cancer categorisation. A pair loss and a triplet loss are proposed to learn embeddings that can measure similarity between images for classification. Furthermore, to eliminate the impact of class imbalance, instead of using the strategy of hard samples mining that intuitively discard some easy samples, we introduce a new loss function to simultaneously punish hard misclassified samples and suppress easy well-classified samples. Tumour area segmentation in whole-slide images is a fundamental step for viable tumour burden estimation, which is of great value for cancer assessment. Vague boundaries and small regions dissociated from viable tumour areas are two main challenges to accurately segment tumour area. We present a structure-aware scale-adaptive feature selection method for efficient and accurate tumour area segmentation. Specifically, based on a segmentation network with a popular encoder-decoder architecture, a scale-adaptive module is proposed to select more robust features to represent the vague, non-rigid boundaries. Furthermore, a structural similarity metric is proposed for better tissue structure awareness to deal with small region segmentation. Detection of cell-level instances in histology images is essential to acquire morphological and numeric clues for cancer assessment. However, multiple reasons such as morphological variations of nuclei or cells make it a challenging task where conventional object detection methods cannot obtain satisfactory performance in many cases. We propose similarity-based region proposal networks for nuclei and cells detection in histology images. In particular, a customized convolution layer termed as embedding layer is designed for network building. The embedding layer is then added on to modify the region proposal networks, which enables the networks to learn discriminative features based on similarity learning.
Authors
SUN, YCollections
- Theses [4235]