Machine Learning Assisted Digital Pathology

Histopathological tissue samples contain a vast amount of information on underlying biological mechanisms that contribute to disease manifestation and progression. Therefore, diagnosis from histopathological tissue samples has been the gold standard for decades. However, traditional histopathological assessment is a laborious task and prone to human errors, thereby leading to misdiagnosis or delayed diagnosis. The development of whole slide scanners for digitization of tissue glass slides has initiated the transition to a fully digital pathology workflow that allows scanning, interpretation, and management of digital tissue slides. These advances have been the cornerstone for developing intelligent algorithms and automated computational approaches for histopathological assessment and clinical diagnostics.

Machine learning is a subcategory of artificial intelligence and can be defined as a process of learning from data. In image analysis tasks, the raw pixel values are transformed into quantitative feature representations. Based on the image data representation, a machine learning model learns a set of rules that can be used to extract meaningful information and knowledge. Over the years, the field of machine learning based image analysis has developed from manually handcrafting complex features to the recent revolution of deep learning and convolutional neural networks. Histopathological assessment can benefit greatly from the ability of machine learning models to discover patterns and connections from the data. Therefore, machine learning holds great promise to improve the accuracy, reproducibility, and efficiency of clinical diagnostics in the field of digital pathology.

This thesis is focused on developing machine learning based methods for assisting in the process of histopathological assessment, which is a significant step in clinical diagnostics as well as in preclinical studies. The studies presented in this thesis show the effectiveness of feature engineering and machine learning in histopathological assessment related tasks, such as; tissue characterisation, metastasis detection, epithelial tissue detection, and nuclei detection. Moreover, the studies presented in this thesis address the key challenges related to variation presented in histopathological data as well as the generalisation problem that need to be considered in order to integrate machine learning approaches into clinical practice. Overall, these studies have demonstrated the potential of machine learning for bringing standardization and reproducibility to the process of histopathological assessment.