Bilkent University
Department of Computer Engineering


Deep Learning for Histopathological Image Modeling


Can Taylan Sarı
PhD Student
Computer Engineering Department
Bilkent University

Digital pathology is mainly focused on developing a method of segmenting a given histopathological image into homogeneous areas and classifying them with a specific category. These segmentation and classification methods work on mathematical notations. In studies in the literature, typically human hand-designed features (eg, graph-based or textual features) have been extracted on image pixels or image segments in order to identify these notations. On the other hand, the design of human-handed features may require the use of application-dependent field knowledge, and sometimes does not reflect exactly the image characteristics, depending on the attributes to be defined. Deep Learning algorithms, which are formed by combining linear and non-linear transformations of the data, are intended to obtain successive, and eventually more useful, hierarchically higher representations of the data. Representations at different levels (corresponding transformations) are automatically learned from image pixels. Thus, the obtained hierarchical representations (attributes) are learned directly on image pixels without using field knowledge. A learning process conducted in this way has the potential of reflecting the image characteristics better. Deep learning algorithms are one of the most popular research areas in recent years with superior performance in different pattern recognition and computer vision problems. On the other hand, the use of deep learning algorithms on medical and histopathological images is limited to a few studies, especially in the area of ??cell segmentation. The aim of this doctoral dissertation is to define innovative representations for histopathologic images taken from colon tissue by using deep learning; and to develop effective classification and segmentation methods using these representations. Accordingly, we continue our studies by focusing on deep belief networks and stacked autoencoders besides convolutional neural networks.


DATE: 05 December, 2016, Monday @ 16:00