Unsupervised image segmentation methods
Thesis title in Czech: | Neřízené metody segmentace obrazu |
---|---|
Thesis title in English: | Unsupervised image segmentation methods |
Key words: | Neřízené (unsupervised) metody|Segmentace obrazu |
English key words: | Unsupervised methods|Image segmentation |
Academic year of topic announcement: | 2022/2023 |
Thesis type: | dissertation |
Thesis language: | angličtina |
Department: | Department of Software and Computer Science Education (32-KSVI) |
Supervisor: | doc. RNDr. Elena Šikudová, Ph.D. |
Author: | hidden - assigned and confirmed by the Study Dept. |
Date of registration: | 16.08.2022 |
Date of assignment: | 16.08.2022 |
Confirmed by Study dept. on: | 27.09.2022 |
Guidelines |
Annotated data is extremely difficult to obtain, requiring extensive annotation or skilled annotation by domain experts, which is often a huge problem in the medical domain. In the absence of ground truths, more focus needs to be applied to exploring unsupervised learning approaches.
Unsupervised segmentation methods are more generally applicable and more robust to atypical or unseen situations. In addition, the results of such methods are potential starting points for manual segmentation, thereby expediting the creation of training datasets. Since even an unsupervised segmentation algorithm may be effective in segmenting some but not all classes of images, it is important to explore new methodologies to expand the suitable options for unsupervised segmentation of various classes of medical images. Image segmentation can be viewed as a special case of the ML clustering methods, where the difference between them is that clustering usually ignores pixel layout, while segmentation relies heavily on spatial cues and constraints. We will investigate both approaches, ML clustering methods as well as traditional computer vision methods of image segmentation including Spectral Clustering, K-Means Clustering, DBSCAN, CNN, Autoencoders, Mean Shift, Image Morphology, Active Contours, Watershed, Region Merging. In segmenting retinal fundus images, the unsupervised methods are successfully used in blood vessel segmentation, while in anomaly detection the results are not competitive yet. We will focus on improving the performance of the unsupervised methods by investigating the statistical properties of the images, image-preprocessing methods, and feature selection and engineering. |
References |
Richard Szeliski. Computer Vision - Algorithms and Applications, Second Edition. Texts in Computer Science. Springer, 2022 |