Highlighting dissimilarity in medical images using hierarchical clustering based segmentation (HCS).

Tissue abnormality in a medical image is usually related to a dissimilar part of an otherwise homogeneous image. The dissimilarity may be subtle or strong depending on the medical modality and the type of abnormal tissue. Dissimilarity within an otherwise homogeneous area of an image may not always be due to tissue abnormality. It might be due to image noise or due to variability within the same tissue type. Given this situation it is almost impossible to design and implement a generic segmentation process that will consistently give a single appropriate solution under all conditions. Hence a dissimilarity highlighting process that yields a hierarchy of segmentation results is more useful. This would benefit from high level human interaction to select the appropriate image segmentation for a particular application, because one of the capabilities of the human vision process when visualising images is its ability to visualise them at different levels of details.The purpose of this thesis is to design and implement a segmentation procedure to resemble the capability of the human vision system's ability to generate multiple solutions of varying resolutions. To this end, the main objectives for this study were: (i) to design a segmentation process that would be unsupervised and completely data driven. (ii) to design a segmentation process that would automatically and consistently generate a hierarchy of segmentation results. In order to achieve these objectives a hierarchical clustering based segmentation (HCS) process was designed and implemented. The developed HCS process partitioned the images into their constituent regions at hierarchical levels of allowable dissimilarity between the different spatially adjacent or disjoint regions. At any particular level in the hierarchy the segmentation process clustered together all the pixels and/or regions that had dissimilarity among them which was less than or equal to the dissimilarity allowed for that level. The clustering process was designed in such a way that the merging of the clusters did not depend on the order in which the clusters were evaluated.The HCS process developed was used to process images of different medical modalities and the results obtained are summarised below: (i) It was successfully used to highlight hard to visualise stroke affected areas in T2 weighted MR images confirmed by the diffusion weighted scans of the same areas of the brain. (ii) It was used to highlight dissimilarities in the MRI, CT and ultrasound images and the results were validated by the radiologists. It processed medical image data and consistently produced a hierarchy of segmentation results but did not give a diagnosis. This was left for the experts to make use of the results and incorporate these with their own knowledge to arrive upon a diagnosis. Thus the process acts as an effective computer aided detection (CAD) tool.The unique features of the designed and implemented HCS process are: (i) The segmentation process is unsupervised, completely data driven and can be applied to any medical modality, with equal success, without any prior information about the image data(ii) The merging routines can evaluate and merge spatially adjacent and disjoint similar regions and consistently give a hierarchy of segmentation results. (iii) The designed merging process can yield crisp border delineation between the regions.