Automated analysis of necrosis and steatosis in histological images : Practical solutions for coping with heterogeneity and variability

Abstract

Pathological examination of histological tissue sections is essential for the diagnosis of many life-threatening diseases. Demographic change and the growing importance of precision medicine require pathology to become more efficient, reproducible and quantitative. Automated histological image analysis is an important tool to meet these demands. This thesis is based on five research papers that consider specific problems in histological image analysis. The problems are related either to the quantification of necrosis or to the quantification of steatosis in histological sections of liver tissue. Both are typical applications in which tissue structures or cellular structures must be identified and quantitatively analyzed. In this context, the papers address important general challenges in histological image analysis and present broadly applicable solutions. One challenge is spatial heterogeneity of tissue properties, which can make their quantification sensitive to tissue sampling and image analysis errors. As a solution, the papers present novel scores that enable reliable measurement of heterogeneously distributed tissue properties. Another challenge is the huge variability of histological images, which can make machine learning-based analysis methods require large amounts of training data to work robustly. As a solution, the papers show how interactive training can produce accurate results with little training effort. Finally, a practical challenge is achieving a good trade-off between accuracy, efficiency, and simplicity. In this regard, the papers describe pragmatic approaches to enable accurate and fast analysis of gigapixel images on standard computers.