Automated classification of HEp-2 cell images is crucial for fast and accurate detection of autoimmune diseases. Recent competitions resulted in high classification rates on publicly available datasets. However, performance on low-resolution HEp-2 images typically lagged behind that of high-resolution images due to the blurring and sub-sampling of fine cellular details. Direct interpolation of low-resolution images couldn't fill in the performance gap. We propose a learning-based approach to infer high-resolution features from low-resolution HEp-2 images. Our approach exploits partial least-square (PLS) to linearly project low- and high-resolution HEp-2 image features into a common linear subspace where the features are highly correlated. Experimental and statistical evaluations show significant improvement in classification performance due to our technique in comparison with direct interpolation and canonical correlation analysis techniques.

Analysis of retinal vessels in fundus images provides a valuable tool for characterizing many retinal and systemic diseases. Accurate automatic segmentation of these vessels is usually required as an essential analysis step. In this work, we propose a new formulation of deep Convolutional Neural Networks that allows simple and accurate segmentation of the retinal vessels. A major modification in this work is to reduce the intra-class variance by formulating the problem as a Three-class problem that differentiates: large vessels, small vessels, and background areas. In addition, different sizes of the convolutional kernels have been studied and it was found that a combination of kernels with different sizes achieve the best sensitivity and specificity. The proposed method was tested using DRIVE dataset and it showed superior performance compared to several other state of the art methods. The segmentation sensitivity, specificity and accuracy were found to be 83.97%, 95.62% and 94.56% respectively.

Active shape models (ASM) showed to have potential for segmenting the right ventricle (RV) in cardiac magnetic resonance images (MRIs). Nevertheless, the large variability and complexity of the RV shape do not allow for concisely capturing all possible shape variations among patients and anatomical cross‐sections. Noticeably, the latter increases the number of iterations required to converge to a proper solution and reduces the segmentation accuracy. In this study, the authors propose a new ASM framework that can model the RV shape in short‐axis cardiac MRI images. In this framework, the RV contour is split into two simpler segments, septal (SP) and free wall, whose shape variations are independently modelled using two separate (dual) ASM models. The contour splitting is done at the location of the RV insertion points into the SP wall. Further, instead of using the conventional Procrustes method, the RV contours are aligned using the Bookstein coordinate transformation, which uses the RV insertion points as landmarks to linearly align the RV contours. The results from a dataset of 14 patients show that the proposed framework outperforms the conventional ASM framework and can model complex RV shape variation with more accuracy and in less iteration steps.

Fundus fluorescein angiography (FFA) is a standard screening and diagnosis technique for several retinal diseases. The analysis of FFA images is performed qualitatively by skilled observers, and thus is vulnerable to inter‐ and intra‐observer variability. In this study, the authors present a method for computer‐aided analysis of FFA images. The method is based on generating quantitative colour fluorescein leakage maps (FLM) that mimic the thickness maps generated by the optical coherence tomography (OCT). Results from 64 patients show strong correlation between the FLM and OCT thickness maps ( r = 0.8). The method was found to be reproducible and robust to variability in the image acquisition times.

Cardiac magnetic resonance imaging provides a number of different imaging acquisition types and views of different body cross sections and orientations. A huge amount of images are produced which demand an automatic method for classification based on the visual contents to facilitate diagnosis and searching operations. In this work, we propose a fully automated classification method for classifying cardiac MRI images according to image acquisition type and orientation. Local binary pattern is used to represent the texture differences among the different image types. Edge orientation histogram is used to differentiate the different image orientations. In addition, two similarity measures are applied and compared: log-likelihood and chi-square distance. The chi-square similarity measure showed better results than the log-likelihood. An average accuracy for classifying the image type and orientation using chi-square was respectively 97% and 96%.

Optical flow (OF) tracking of the myocardium contours has a potential in segmenting the myocardium in time sequences of cardiac medical images. Nevertheless, to estimate the displacement field of the contour points, a number of assumptions are required to solve an under‐determined set of optical flow equations. In this work, a new framework is proposed to solve the OF tracking problem using greedy optimisation algorithm. The new framework allows different types of constraints such as motion invariance, shape and topology to be applied in a unified way. The developed methods are applied to a publicly‐available cardiac magnetic resonance imaging dataset containing image sequences for 33 patients. Quantitative evaluation of the results shows high potential of the methods to accurately track and segment the myocardium contours.

Colour fundus images play an important role in diagnosing and screening diabetic macular edema (DME). In rural areas, content-based image retrieval (CBIR) might compensate the lack of expert ophthalmologists. In this work, we present a fully automated CBIR system that retrieves fundus images according to their content (quantity and location) of exudates. First, the macula is divided into three concentric regions whose texture discontinuities are used to represent lesion content of the retina. The image-to-image distance measure gives higher weights to lesions closer to the fovea to reflect the severity of the DME. Retrieval results of the system show precision of 79.2%.

Myocardium rotation and torsion are important indicators of the cardiac function. Currently, tagged Magnetic Resonance Imaging (tMRI) sequences are analyzed to estimate these parameters. Unfortunately, tMRI is not widely used in clinical practice because it prolongs the scanning time and requires sophisticated analysis software. In this work, we present a method for estimating the myocardium rotation from standard cine MRI sequences. The method is based on identifying special features, i.e. landmarks, of the intensity pattern around the myocardium borders at each timeframe. Each set of landmarks is then encoded using a string of characters that can be matched across the different timeframes. String alignment technique is used to match the characters and thus determine the inter-frame motion of each landmark. Preliminary results using dataset of three patients (apical, mid, and basal slices) indicate the potential of the method to estimate the myocardium rotation especially at late timeframes where tMRI fails due to tag fading.

Evaluation of cardiac functions using Strain Encoded (SENC) magnetic resonance (MR) imaging is a powerful tool for imaging the deformation of left and right ventricles. However, automated analysis of SENC images is hindered due to the low signal‐to‐noise ratio SENC images. In this work, the authors propose a method to segment the left and right ventricles myocardium simultaneously in SENC‐MR short‐axis images. In addition, myocardium seed points are automatically selected using skeletonisation algorithm and used as hard constraints for the graph‐cut optimization algorithm. The method is based on a modified formulation of the graph‐cuts energy term. In the new formulation, a signal probabilistic model is used, rather than the image histogram, to capture the characteristics of the blood and tissue signals and include it in the cost function of the graph‐cuts algorithm. The method is applied to SENC datasets for 11 human subjects (five normal and six patients with known myocardial wall motion abnormality). The segmentation results of the proposed method are compared with those resulting from both manual segmentation and the conventional histogram‐based graph‐cuts segmentation algorithm. The results show that the proposed method outperforms the histogram‐based graph‐cuts algorithm especially to segment the thin structure of the right ventricle.

Diagnosis of Diabetic Macular Edema (DME) from Fundus Fluorescein Angiography (FFA) image sequences is a standard clinical practice. Nevertheless, current methods depend on subjective evaluation of the amount of fluorescein leakage in the images which lack reproducibility and require well-trained grader. In this work, we present a method for processing FFA images to generate a fluorescein leakage map that can be used for quantitative evaluation of DME. An essential step in the proposed method is to model the spatial distribution of the image intensity within the macula. This model, which represents the non-leaking structures, is then subtracted from the late timeframe image to enhance the areas of fluorescein leakage. The resulting difference image is then mapped with empirical linear transformation to produce a color Fluorescein Leakage Map (FLM) that can be used for quantitative assessment and detection of DME. The method was applied to 13 image sequences for 13 different patients. The resulting FLM maps were found to be correlated with the thickness maps produced by Optical Coherence Tomography (OCT). The relatively high correlation between the FLM and OCT maps show the potential and of using the developed method for quantitatively assess the DME in FFA image sequences.

Optic Disc (OD) localization is an important pre-processing step that significantly simplifies subsequent segmentation of the OD and other retinal structures. Current OD localization techniques suffer from impractically-high computation times (few minutes per image). In this work, we present a fast technique that requires less than a second to localize the OD. The technique is based upon obtaining two projections of certain image features that encode the x- and y- coordinates of the OD. The resulting 1-D projections are then searched to determine the location of the OD. This avoids searching the 2-D image space and, thus, enhances the speed of the OD localization process. Image features such as retinal vessels orientation and the OD brightness are used in the current method. Four publicly-available databases, including STARE and DRIVE, are used to evaluate the proposed technique. The OD was successfully located in 330 images out of 340 images (97%) with an average computation time of 0.65 s.

Optic Disc (OD) localization is a fundamental step in developing computer-assisted diagnostics. In this work, an ultrafast method to locate the OD in retinal fundus images is presented. The proposed method is based on transforming the localization problem into two 1D problems by projecting the image features onto two perpendicular directions. Image features such as the directionality of the retinal vessels, the brightness and the size of the OD have been used in the current method. Two publicly available databases were used to evaluate the accuracy and the computation time of the proposed technique. The OD was correctly located in 71 out of 81 of the STARE images (87.6%) with computation time of 0.8 seconds per image. The OD was also successfully located in all the 40 images (100%) using the DRIVE database.

In magnetic resonance imaging (MRI), data are collected as spectrum samples. The acquisition time is proportional to the number of the spectrum lines. Therefore, only few lines of the data space may be required in order to track rapid changes of an object. In the current techniques, the missed lines may be zeroed or replaced by the corresponding lines in a reference image, which is acquired a priori for the same anatomical cross-section. However, this always comes at the expense of the spatial-resolution. In this study, we propose an extrapolation iterative algorithm to provide an improved estimate of the missed lines. Additional spatial and spatial-frequency constraints of the reference image are incorporated to enhance the convergence and obtain a better estimate of the initial conditions of the iterations. Results from simulated data verify the theory and indicate that the algorithm may provide better reconstruction in dynamic imaging studies.

In magnetic resonance imaging, some schemes require slow data acquisition in order to achieve acceptable spatial resolution and signal-to-noise ratio. However, long scanning periods allow patient motion to distort the reconstructed images. In cases of planar rigid body motion, rotational and translational transformations can be assumed to model the patient motion. In this study, we develop a novel technique for estimating the parameters of these transformations. It is based on fixing two markers to the patient and using a post-processing algorithm to estimate the location of these two markers during the scan time. The key factor in this algorithm is to incorporate a priori information about the geometric shape of the markers to achieve accurate estimation. Computer simulation results confirmed the ability of the algorithm to estimate the unknown parameters.