Segmentation of highly corrupted medical images

Abstract

Image segmentation is generally not straightforward and is still a fundamental problem in many fields, while there have been many different approaches to solve it, to date there is no definitive answer. While there is no generalised solution which can match human cognition, we can attempt to design algorithms which can work well for specific applications. In the case of medical imaging, there are a myriad of problems facing a successful segmentation algorithm depending on which particular problem it is being applied to. One such subset of this field is images with intensity inhomogeneities and corruptions. Although this is not a problem limited to the medical imaging, it does appear in many situations. Another issue often faced in medical imaging when trying to create a fully automated system, is a lack of a reliable ground truth or "gold standard" to train the model. Moreover, even if one does exist there can often be some disagreement between different medical experts over the its validity. Unsupervised learning is less affected by this paradoxical dilemma, however, validating the results of these methods is slightly tricky and the evaluation of how well any such system is performing is not straight forward. The focus of this research is on how to correctly segment and classify pixels in images which contain such inhomogeneities, as well as their classification for which-ever particular problem is being tackled. Three distinct applications are given, that of Multicolour-fluorescence in situ hybridization (M-FISH) imaging, retina blood vessel images and Magnetic resonance imaging (MRI). For each of these applications four methods will be presented. First an investigation into how well an adaptive Kernelised Fuzzy C-Means algorithm (FCM) can be used to segment pixels and also how the resultant output can be used intuitively in a Bayesian classifier. Specifically the research presented shows the results of the segmentation of M-FISH image data using a Kernel-based modifed adaptive FCM algorithm; the kernel used was the Radial basis function exclusively. The research then shows how the responsibility values from the FCM algorithm can be used to form a set of probabilities for use in a Bayesian classifier, achieved by assuming that the intensity values of the constituent images are strongly correlated. To date the proposed algorithm has shown a vast improvement on the standard segmentation method. Secondly, a method combining both pixel-wise and geometric-based segmentation methods could be used together to inform one another. The outcome was that by fusing together these two methods it was possible to get a better result than either used alone. Specifically looking at images at intensity inhomogeneities, some methods have tried to incorporate handling these into both region-based and geometric based, from this study it can be seen that by combining the two methods together they show a better result than individually. Finally, a method which fuses the benefits of both of the adaptive fuzzy algorithms and Level-sets if presented. Using the developments from both of the previous methods, a conditional energy formulation for a Level-set curve evolution is made by leveraging the uncertainties calculated from an adaptive FCM algorithm. This method shows a strong capability of overcoming the corruptions within M­FISH images with segmentation performance much higher than those it is compared with. The proposed methods are verified using experimentation on public datasets and synthetic images. From the results it was seen that the proposed methods in this study can improve segmentation performance on small datasets of medical images which contain both inhomogeneities and corruptions which shows it has potential for use in real world applications of health informatics in the future.