School of Computing Graduate Theseshttp://hdl.handle.net/1974/7922017-09-23T01:57:55Z2017-09-23T01:57:55ZFractional Labelmap Representation of Anatomical StructuresSunderland, Kylehttp://hdl.handle.net/1974/226772017-09-16T04:46:19ZFractional Labelmap Representation of Anatomical Structures
Sunderland, Kyle
INTRODUCTION: In medical imaging software, structures are often represented using image volumes known as labelmaps. Conversion of structures to labelmap representations results in a loss of information which impacts medical imaging algorithms, including those in radiation therapy (RT) treatment planning. RT treatment planning systems are used to optimize radiation delivery to structures, which are represented as planar contours. Errors from conversion affect metrics such as dose volume histograms (DVHs) that are the primary metric for RT plan optimization. The goal of this thesis is to develop fractional labelmaps as a structure representation that preserves more structural information and reduces conversion errors. METHODS: The effect of voxelization on treatment planning metrics was tested by comparing DVHs calculated using varying voxel sizes. An algorithm was implemented that triangulates planar contours to closed surface mesh and handles features such as branching, keyhole contours, and end-capping. Closed surfaces were converted into fractional labelmap representations, where each voxel represents occupancy between 0% and 100%. Existing segmentation methods were modified to allow fractional labelmaps to be edited directly. RESULTS: Algorithms were implemented in the open-source medical imaging platform 3D Slicer, and the SlicerRT toolkit. Voxel size was found to affect DVH accuracy for structures with small features or in regions with a high dose gradient. The planar contour to closed surface triangulation algorithm produced qualitatively good surfaces. Fractional labelmaps from closed surfaces were found to be up to 19.1% better at representing structure volume than binary labelmaps, with an average improvement of 6.8%. DVHs calculated from fractional labelmaps were found to be more accurate than DVHs calculated using binary labelmaps. Fractional segmentation methods were found to create good quality segmentations. CONCLUSION: Labelmap voxel size was found to be a contributing factor for DVH accuracy. Accurate conversion algorithms were implemented for planar contour to closed surface and closed surface to fractional labelmap conversions. When used for structure representation and DVH calculation, fractional labelmaps were more accurate than binary labelmaps at the same resolution. Fractional labelmaps were found to be an effective tool for structure representation in radiotherapy that could be expanded to other use cases.
Context Sensitive and Secure Parser Generation for Deep Packet Inspection of Binary ProtocolsEl Shakankiry, Alihttp://hdl.handle.net/1974/220402017-09-15T13:19:12ZContext Sensitive and Secure Parser Generation for Deep Packet Inspection of Binary Protocols
El Shakankiry, Ali
Network protocol parsers constantly dissect a large number of network data to place into internal data structures for further processing by traffic analysis systems. Many network protocol parsers are hand-written for performance reasons, and lack the security required to run on mission-critical networks. We propose an approach that automatically generates custom protocol parsers to process network traffic to be used as part of an Intrusion Detection System. The user is provided a specification language in which they can define the protocols they need to analyse. This thesis looks at command and control/industrial control networks that are characterized by a limited number of known protocols. We present a robust, secure, and high-performing solution that deals with the issues that have only partially been addressed in this domain.
Distances Between Languages: Algorithms and Descriptional ComplexityNg, Timothyhttp://hdl.handle.net/1974/220182017-09-15T13:18:58ZDistances Between Languages: Algorithms and Descriptional Complexity
Ng, Timothy
We consider the descriptional complexity of neighbourhoods of regular languages and the computational complexity of computing the distance between languages. Distance measures are defined on words to describe their similarity. These measures can be extended to languages, which are sets of words. We consider several different notions of distance measures on languages.
The neighbourhood of a language L is the set of words within some fixed distance of a word in L. We show that we can construct weighted finite automata for neighbourhoods of regular languages with respect to additive quasi-distances. We consider the state complexity of the neighbourhoods and the state complexity of approximate pattern matching. We establish a tight state complexity lower bound for the edit distance neighbourhood of a language recognized by a deterministic finite automaton.
We consider the deterministic and nondeterministic state complexity of prefix, suffix, and subword distance neighbourhoods. We establish tight deterministic state complexity bounds for the prefix and suffix distance neighbourhoods of regular languages and of finite languages. For prefix distance neighbourhoods, we consider prefix-convex subclasses of regular languages. We also establish the state complexity of suffix neighbourhoods of suffix-closed languages. We give an upper bound for the state complexity of the subword distance neighbourhood of a regular language.
Based on the edit distance algorithms of Mohri and Han and co-authors, we give algorithms for computing the prefix distance between two regular languages, a context-free language and a regular language, and two visibly pushdown languages. We extend these algorithms to compute the inner prefix distance of regular languages and visibly pushdown languages.
We give algorithms to compute the relative prefix distance of two regular languages. For a given integer k, we show that for a deterministic context-free language and a regular language or for two visibly pushdown languages, it is decidable whether their relative prefix distance is at most k. This is very close to the borderline between decidable and undecidable since, on the other hand, we show that deciding the same question for a deterministic context-free language and a visibly pushdown language is undecidable.
Discovery of Patterns in Simulink Systemsde la Parra, Franciscohttp://hdl.handle.net/1974/201022017-08-10T12:20:27ZDiscovery of Patterns in Simulink Systems
de la Parra, Francisco
Model-Driven Engineering tools and techniques have become mainstream aids in the development
of embedded cyber-physical systems. The large-scale use of visual models is a powerful
paradigm for designing reliable and maintainable systems assembled with complex hardware, software
and mechanical elements. Modular design approaches can readily capture the requirements
for embedded systems into executable and evolutionary models.
We propose a framework to discover, classify and visualize model-patterns in large repositories of
Simulink models. We design, and implement in a software toolchain, a set of scalable algorithms to
fully traverse and contextually parse the source code of these repositories in order to compute topology
hash functions for the size, name connectivity and type connectivity of each model subsystem,
as well as properties of Stateflow charts. We use these extracted properties to cluster the subsystems
into a set of addressable-by-property classes. We develop an interactive model visualization and
querying facility, embodied by the MoSART user interface, to identify further relations, similarities
and abstractions in the elements of the cluster classes. This tool provides a configurable environment
for inferring structural model-patterns and elaborating on semantic model-patterns with the assistance
of skilled domain knowledge. We propose a query system, based on regular expressions that
match values of previously computed subsystems properties, as an effective method to iteratively
refine Simulink model-patterns abstractions into accumulated modeling knowledge.
Our implementation works cooperatively with a 3-phase iterative model-pattern discovery approach
that we conceptualize as 1) model clustering: repository traversal and subsystem-topology
evaluation for clustering subsystems, 2) primary model-pattern inference: visualization, querybased
search and clustering for identifying basic semantic patterns, 3) model-pattern refinement:
continuous refinement of basic patterns into specialized groups for application in targeted use cases.
We effectively support the iterative nature of this approach because our repository traversal, property
evaluation and clustering algorithms have low time-complexity, that is, between a lower-bound(n) and
O(n lg n)(n = total lines of source code in a repository), and our interactive tool provides a
complete-coverage view of a model repository.