The morphological analysis of blood smear slides by haematologists or haematopathologists is one of the diagnostic procedures available to evaluate the presence of acute leukaemia. This operation is a complex and costly process, and often lacks standardized accuracy owing to a variety of factors, including insufficient expertise and operator fatigue. This research proposes an intelligent decision support system for automatic detection of acute lymphoblastic leukaemia (ALL) using microscopic blood smear images to overcome the above barrier. The work has four main key stages. (1) Firstly, a modified marker-controlled watershed algorithm integrated with the morphological operations is proposed for the segmentation of the membrane of the lymphocyte and lymphoblast cell images. The aim of this stage is to isolate a lymphocyte/lymphoblast cell membrane from touching and overlapping of red blood cells, platelets and artefacts of the microscopic peripheral blood smear sub-images. (2) Secondly, a novel clustering algorithm with stimulating discriminant measure (SDM) of both within- and between-cluster scatter variances is proposed to produce robust segmentation of the nucleus and cytoplasm of lymphocytic cell membranes. The SDM measures are used in conjunction with Genetic Algorithm for the clustering of nucleus, cytoplasm, and background regions. (3) Thirdly, a total of eighty features consisting of shape, texture, and colour information from the nucleus and cytoplasm of the identified lymphocyte/lymphoblast images are extracted. (4) Finally, the proposed feature optimisation algorithm, namely a variant of Bare-Bones Particle Swarm Optimisation (BBPSO), is presented to identify the most significant discriminative characteristics of the nucleus and cytoplasm segmented by the SDM-based clustering algorithm. The proposed BBPSO variant algorithm incorporates Cuckoo Search, Dragonfly Algorithm, BBPSO, and local and global random walk operations of uniform combination, and Lévy flights to diversify the search and mitigate the premature convergence problem of the conventional BBPSO. In addition, it also employs subswarm concepts, self-adaptive parameters, and convergence degree monitoring mechanisms to enable fast convergence. The optimal feature subsets identified by the proposed algorithm are subsequently used for ALL detection and classification. The proposed system achieves the highest classification accuracy of 96.04% and significantly outperforms related meta-heuristic search methods and related research for ALL detection.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:757246 |
| Date | January 2017 |
| Creators | Srisukkham, Worawut |
| Contributors | Zhang, Li |
| Publisher | Northumbria University |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://nrl.northumbria.ac.uk/36140/ |
Page generated in 0.0015 seconds