Abstract

Sparse matrix-sparse/dense matrix multiplications, spgemm and csrmm, respectively, among other applications find usage in various matrix formulations of graph problems. Considering the difficulties in executing graph problems and the duality between graphs and matrices, computations such as spgemm and csrmm have recently caught the attention of HPC community. These computations pose challenges such as load balancing, irregular nature of the computation, and difficulty in predicting the output size. It is even more challenging when combined with the GPU architectural constraints such as memory accesses, limited shared memory, strict SIMD and thread execution. To address these challenges on a GPU, we evaluate three possible variations of matrix multiplication (Row-Column, Column-Row, Row-Row) and perform suitable optimizations targeted at sparse matrices. Our experiments indicate that the Row-Row formulation, which mostly outperforms the other formulations, is 3.5x faster on average compared to an optimized multi-core implementation in the Intel MKL library. We extend the Row-Row formulation to a CPU+GPU hybrid algorithm that simultaneously utilizes the CPU also. In this direction, we present heuristics to find the right amount of work division between the CPU and the GPU. Our hybrid row-row formulation of the spgemm operation performs 5.5x faster on average when compared to the optimized multi-core implementation in the Intel MKL library. Our experience indicates that it is difficult to identify right amount of work division between the CPU and the GPU. We therefore investigate a subclass of sparse matrices, band matrices, and present an analytical method to identify a good work division when multiplying two band matrices. Our GPU csrmm operation performs 2.5x faster on average when compared to a corresponding implementation in the cusparse library, which outperforms the Intel MKL library implementation.

Abstract

Accurate segmentation of the cup region from retinal images is needed to derive relevant measurements for glaucoma assessment. A novel, depth discontinuity (in the retinal surface)-based approach to estimate the cup boundary is proposed in this paper. The proposed approach shifts focus from the cup region used by existing approaches to cup boundary. The given set of images, acquired sequentially, are related via a relative motion model and the depth discontinuity at the cup boundary is determined from cues such as motion boundary and partial occlusion. The information encoded by these cues is used to approximate the cup boundary with a set of best-fitting circles. The final boundary is found by considering points on these circles at different sectors using a confidence measure. Four different kinds of data sets ranging from synthetic to real image pairs, covering different multiview scenarios, have been used to evaluate the proposed method. The proposed method was found to yield an error reduction of 16% for cup-to-disk vertical diameter ratio (CDR) and 13% for cup-to-disk area ratio (CAR) estimation, over an existing monocular image-based cup segmentation method. The error reduction increased to 33% in CDR and 18% in CAR with the addition of a third view (image) which indicates the potential of the proposed approach.

Abstract

Retinopathy of prematurity (ROP) is a vascular disease in premature infants. This is characterized by abnormal vessel growth and subsequent fibrosis in the peripheral retina. The prognosis of ROP relies on information on the presence of abnormal growth and their location. Diagnosis is based on a series of images obtained with a wide field of view camera (such as RetCam), to capture the complete retina. In this paper, we present a novel and efficient hierarchal mosaicing algorithm, for neonatal images of varying quality. We employ a vessel-based quality metric and exploit the knowledge of the retinal structure to automatically select a subset of images from a given set, to construct a good quality mosaic. Such mosaics can aid the assessment of ROP by providing a comprehensive and complete view of the entire retina. The hierarchal approach underlying the method makes it possible to complete the mosaicing task in close to real time. The proposed method has been tested on 14 sets of data with each set consisting of 6-35 retinal images acquired using RetCam and the generated mosaics are found to be of good quality as validated by a clinical expert.

Abstract

The task of detecting peri-papillary indicators associated with the glaucoma is challenging due to the high degree of intra and inter-image variations commonly seen in colour retinal images. The existing approaches based on direct modeling of the region of interest failto handle such image variations which compromises detection accuracy.In this paper, a novel detection strategy is proposed which exploits the saliency property associated with these indicators. The region of interest is modeled as a region substantially different from the adjacent image regions. This dissimilarity information is derived at the feature level, be-tween the target and its adjacent regions. Based on the proposed strategy,two novel methods are presented for the detection of peri-papillary atrophy and RNFL defect from colour retinal images. Two different datasets have been used to evaluate the performance of developed solutions. The obtained results are encouraging and establish the strength of the pro-posed strategy in handling high degree of image variations. The preliminary results and comparative evaluation with direct modeling strategy show viability of proposed strategy to be used in the glaucoma assessment task.

Abstract

We designed, developed and evaluated an Augmentative and Alternative Communication (AAC) system,AutVisComm,for children with autism that can run on smart phones and tablets. An iterative de-sign and development process was followed, where the prototypes were developed in close collaboration with the user group, and the usability testing was gradually expanded to larger groups. In the last evaluation stage described here, twenty-four children with autism used AutVisCommto learn to request the desired object. We measured their learning rates and correlated them with their behavior traits (as observed by their teachers) like joint attention, symbolic processing and imitation.We found that their ability for symbolic processing did not correlate with the learning rate, but their ability for joint attention did. This suggests that this system (and this class of AACs) helps to compensate for a lack of symbolic processing, but not for a lack of joint-attention mechanism.

Abstract

Colorization is the process of adding colors to gray scale images. This is done to restore or enhance old films or photographs. Most of the techniques for the colorization of an image or video require manual designation of the locations to be colored and the colors themselves which is an expensive and time consuming process. In this paper, we propose a fast but effective fully automated technique for coloring the gray scale image sequences. We define a notion of a most informative frame which is to be coloured manually and exploit the motion field between frames for propagation of the colors to the remaining frames. The proposed technique attempts to provide a method to minimize the amount of labour required for colorization and to decrease the computational cost of this task. The most informative frame is one which has almost all the objects present in that scene. The motion field estimation is based on optical flow. A final refinement step uses similarity based colour filling. Extensive testing of the proposed technique on a large set of videos from movies and animation confirms that it is efficient and effective without any loss of quality.

Abstract

Detection of abnormalities from medical images is of key interest in developing computer-aided diagnostic tools. In this paper, we observe the key challenges for representation and feature extraction schemes to be met for detection of abnormalities by learning normal cases. We introduce an image representation, motivated by the effect of motion on perception of structures. This representation is based on a set of patterns called generalized moment patterns (GMP) generated via induced motion over regions of interest, for learning normal. The proposed GMP has been utilized to develop a scheme for addressing two well-known problems: lesion classification in mammograms and detection of macular edema in color fundus images. The strengths of this scheme are that it does not require anylesion-level segmentation and relies largely on normal images for training which is attractive for developing screening tools. The proposed scheme has been assessed on two public domain datasets,namely, MIAS and MESSIDOR. A comparison against the performance of state of the art method sindicates the proposed scheme to be superior

Abstract

Diffusion Weighted MR Imaging (DWI) is routinely used for early detection of cerebral ischemic stroke.DWI with higher b-values (b=2000) provide improved sensitivity, higher conspicuity and reduced artifacts and thus improve the detectability of smallest infarcts than conventional DWI (b=1000). We propose a novel framework for accurately detecting stroke regions by combining information from multiple sources:b2000,b1000 data and the apparent diffusion coefficient map.The detected lesions are finally segmented using an active contour approach. The proposed method was tested on 41 datasets acquired with different protocols. A comparison of our method with a leading method [3] validates the effectiveness of our approach. The mediandice coefficient, sensitivity and specificity for stroke segmentation were 0.84, 87.07% and 99.90% respectively.The strength of the proposed method is its ability to capture (and accurately segment) the small (and large) lesions in the data which are often missed by segmentation methods operating on a single b-value data.

Abstract

This paper addresses the data corruption that occurs due to patient motion during a scan which is particularly a problem in perfusion weighted MRI due to long scan times. Motion correction is typically the rate-limiting step in processing as each volume has to be registered to a reference volume. This is compounded by the dynamically varying contrast in the volume series due to passage of an injected contrast agent. We propose an efficient two stage motion correction method, consisting of motion detection and a 2-pass registration method for aligning the motion-corrupted volumes. A 2D block-wise phase correlation in central slices is used for the first stage. Alignment employs a strategy which is sensitive to the status of the bolus in the volume and is based on gamma-variate function fitting for intensity correction to handle dynamic contrast in DSC-MRI. Evaluation of the approach shows that it is fast and accurate.

Abstract

Several computational visual saliency models have been pro-posed in the context of viewing natural scenes. We aim to investigate the relevance of computational saliency models in medical images in the context of abnormality detection.We report on two studies aimed at understanding the role of visual saliency in medical images. Diffuse lesions in ChestX-Ray images, which are characteristic of Pneumoconiosis and high contrast lesions such as ‘Hard Exudates’ in retinal images were chosen for the study. These approximately correspond to conjunctive and disjunctive targets in a visual search task. Saliency maps were computed using three popular models namely Itti-Koch [7], GBVS [3] and SR [4].The obtained maps were evaluated against gaze maps and ground truth from medical experts. Our results show that GBVS is seen to perform the best(Mdn. ROC area= 0.77) for chest X-Ray images while SRperforms the best (ROC area= 0.73) for retinal images,thus asserting that searching for conjunctive targets calls for a more local examination of an image while disjunctive targets call for a global examination. Based on the results of the above study, we propose extensions for the two best performing models. The first extension makes use of top down knowledge such as lung segmentation. This is shown to improve the performance of GBVS to some extent. Inthe second case the extension is by way of including multi-scale information. This is shown to significantly (by 28.76%)improve abnormality detection. The key insight from these studies is that bottom saliency continues to play a predominant role in examining medical images.