Medizinische Universität Graz Austria/Österreich - Forschungsportal - Medical University of Graz
Gewählte Publikation:
Kainz, P.
Automated Histopathological Image Analysis of Human Bone Marrow Tissue using Supervised Machine Learning
Doktoratsstudium der Medizinischen Wissenschaft; Humanmedizin; [ Dissertation ] Graz Medical University; 2016. pp. 204
[OPEN ACCESS]
FullText
- Autor*innen der Med Uni Graz:
- Betreuer*innen:
- Ahammer Helmut
- Altmetrics:
- Abstract:
- Classification of cell types in context of the architecture in tissue specimen is the basis of diagnostic pathology. Decisions for comprehensive investigations rely on a valid interpretation of tissue morphology. Especially visual examination of bone marrow cells in gigapixel histopathological images of biopsy sections consumes a considerable amount of time, where both intra- and inter-observer variability can be remarkable. This thesis proposes and evaluates approaches based on supervised machine learning for the automated localization and classification of bone marrow cells and their maturity. Their main advantage is that they work on raw images without relying on segmentation or manual feature extraction. A new method termed proximity score regression is introduced, where employing the Random Forest (RF) algorithm enables an easy implementation. For each image location, a non-linear monotonous function of the distance to the closest cell nucleus center is predicted. Cell centers can then be identified by revealing locally maximal locations in the proximity score map. On five challenging datasets, the proposed approach outperforms current state-of-the-art methods in terms of detection reliability, spatial localization accuracy, and speed. To classify maturation stages, a rotation-invariant classification scheme for multi-class Echo State Networks (ESNs) is proposed. Based on representing 2D single-cell image patches as temporal sequence of rotations, ESNs robustly recognize cells of arbitrary orientations by taking advantage of their short-term memory capacity. A comparison to a standard RF classifier is provided and discussed. This thesis provides evidence that the application of supervised machine learning facilitates reliable image analysis systems, characterized by a predictable error. Driven by the key requirement of having reliable ground truth data, a web-application is presented that enables the conduction of controlled inter-observer reliability studies. While the presented results look promising for computer-aided diagnosis, an assessment of the agreement among algorithms and human observers must be studied in future work.