Department of Computer Engineering
CS 590 SEMINAR
Multitask Learning Approaches for Medical Image Analysis with Applications on Cell Detection
Gözde Nur Güneşli
Computer Engineering Department
Most of the problems in machine learning are, by design, single task learning problems where one type of target is the output for one specific input type. On the other hand, human learning process incorporates knowledge of different areas to learn a task. Considering this, the success of many tasks in the literature can be improved by providing information for different objectives. Cell detection is one of the problems which is usually stated as a single task problem in the literature. The crucial step of this task is to correctly detect cell locations whose difficulty lies along a wide range, from easy to very challenging, depending on the visual characteristics of the cells. This step becomes difficult when the cells appear in varying colors, irregular shapes and overlayers. We propose to approach its learning as a multitask problem which contains both classification and regression tasks. In this study, we present different architectures which address cell detection task as a problem where multiple objective functions are used simultaneously and we compare their results obtained on images acquired by inverted microscopy.
DATE: 05 November, 2018, Monday, CS590 & CS690 presentations begin at @ 15:40