Computer Graphics and
Medical Image Analysis

One of the primary areas of research of this group is Whole Slide Image (WSI) analysis using machine learning algorithms. We have developed several innovative methods for extracting diagnostically relevant features and classifying virtual slides based on tissue/tumour characteristics. Image analysis algorithms have been developed for automated segmentation of regions of interest and quantification of features related to the distribution of intensity, stain, texture and nuclei within a region. Datasets containing whole slide images of H&E and IHC stained slides have been used for performance comparison and evaluation.

Texture plays a very important role in the segmentation and representation of features of biomedical images. Our research group has performed a thorough study into different types of multifractal measures that could represent the distribution of intensity values in pixel neighbourhoods. The well-known Holder Exponent forms the basis of two important descriptors: the alpha-image and the multifractal spectra. Several types of rotation invariant texture features can be derived from these two descriptors. Multifractal features have been used along with Local Binary Patterns and Grey-Level Co-occurrence Matrices in several machine learning applications involving biomedical image segmentation and classification.

The development of computerised mammogram interpretation methods is an active research field. Advanced machine learning and image analysis algorithms are currently being developed to extract diagnostically relevant features and derive quantitative measurements for improving radiologists’ workflow. Two important areas of research in this field are microcalcification detection and breast density classification. Detecting micro-calcifications (MCs) in mammograms is a challenging problem due to heterogeneous properties and diverse composition of breast tissues. For breast density classification in mammograms, extracting effective texture features plays an important role in obtaining accurate classification results. This research project focuses on texture descriptors and the application of machine learning algorithms for the above two tasks.

Histomorphological evaluations of breast conserving surgery specimens in the treatment of Ductal Carcinoma In Situ (DCIS) breast cancer patients would benefit from having a three-dimensional visualization of the tumour structure. 3D reconscturctions of surgical resections also provide a visual guide to extent, proximity to margins, tumour size and extent of involvement of the breast tissue by DCIS. The pathological specimens of a resection are typically stored as tissue images of axial sections at 4mm intervals. Image analysis algorithms are used to segment and extract contours of regions of interest, and piecewise cubic splines are used to generate smooth interpolation curves through matched contour points. Contour point correspondence is established using the dynamic time warping (DTW) algorithm.

Accurate segmentation of substantia nigra (SN) and red nucleus (RN) regions in MRI images is a challenging problem because of their small size, unclear boundaries, morphometric variability and similar intensity profiles with the adjacent structures on standard T1- and, to a lesser extent, T2-weighted MRI. Accurate and precise identification of the structures is key to understanding the processes underlying progression of Parkinson’s disease (PD). This research project involves the development of image analysis algorithms to segment SN and RN from quantitative susceptibility mapping (QSM) MRI. Using Bayesian regression models, we compare QSM values between PD and control groups, and investigate relationships with global cognitive ability and motor severity in PD.

Motion capture data obtained from video sequences find applications in several domains of character animation. This paper considers the problem of estimating three-dimensional motion of human actors from single view video sequences with stationary background. A novel framework based on state-of-the-art Convolutional Neural Networks for object detection is developed. We use a stacked hourglass network to estimate the positions of 2D anatomical landmarks. The projected matching pursuit algorithm is used to recover 3D joint positions from 2D landmarks. Using the proposed method, a quantitative analysis of errors was performed using the Human3.6M dataset.

Ultrasound imaging is widely used for clinical diagnosis owing to several desirable characteristics that minimize health risks such as non-invasiveness and absence of any form of ionizing radiation. In spite of having many advantages in ultrasound imaging, there are certain drawbacks like the speckle noise, low contrast and other artifacts which degrade the visual quality of the image. This research project focuses on algorithms for speckle noise reduction and feature enhancement in ultrasound images and videos.

Vertex skinning is a vital component in modern character animation pipelines.Advanced skinning methods such as example-based skinning and physics-based skinning often require significant computational resources to generate more realistic deformation, while traditional skinning techniques are considerably more suitable for modern GPU rendering and ideal for real-time interactive applications. Traditional vertex skinning methods suffer from large angle rotation artefacts known as the'collapsing-elbow' effect and the 'candy-wrapper effect'. This research project looks at improved skinning methods based on dual quaternions that produce realistic joint angle transformations without the above artefacts.

In the domain of stylistic painterly non-photorealistic (NPR) rendering systems, methods for the characterization of artistic style is for capturing, representing, and remapping a particular artistic style to an input image. Every digitized painting can be seen as a composition of two components: the style and the content. Artistic style characterization process extracts the style component of digitized paintings as a set of features. The features are then used by the NPR system as a heuristic in the painterly rendering process. Methods for iterative brush region extraction, and texture boundary detection are developed and used for extracting shape and texture features of brush strokes. These features are then used as inputs for the artistic style classification algorithm.

Orthogonal moment functions based on Tchebichef polynomials have found several applications in the _eld of image analysis because of their superior feature representation capabilities. Local features represented by such moments could also be used in the design of efficient texture descriptors. In this research project, we have constructed a novel texture descriptor from orthonormal Tchebichef moments evaluated on 5x5 neighborhoods of pixels, and encoded the texture information as a Lehmer code that represents the relative strengths of the evaluated moments. The encoding scheme provides a byte value for each pixel, and generates a gray-level LTM-image. The histogram of the LTM-image is then used as the texture descriptor for image segmentation and classication.

The research group is involved in the development of several computer graphics algorithms for real-time rendering applications. These algorithms leverage the computational power and versatality of shader stages of the OpenGL-4 programmable pipeline. Advanced terrain rendering algorithms capable of rendering highly detailed terrain geometries with dynamic levels of detail and several types of terrain surface features have been developed. Realtime rendering algorithms have also been developed for tree modelling, particle systems, non-photorealistic rendering, character animation using inverse kinematics, and animation retargetting.

In this research, we consider the problem of extracting shape contours from High Resolution Computed Tomography (HRCT) and Magnetic Resonance Imaging (MRI) stacks and using them to construct a three-dimensional mesh surface of the underlying geometry to a high level of detail. While many reconstruction algorithms adopt volumetric approaches and ray casting methods, we propose a novel algorithm for automatic segmentation of large volume sets, and a contour-based construction of a mesh representation that could be used in any rendering application or combined with larger meshes of anatomical parts. Contour based reconstruction also allows parallel computation of the surface segments.