Compressed sensing MRI based on image decomposition model and group sparsity.

The image representation plays an important role in compressed sensing magnetic resonance imaging (CSMRI). However, how to adaptive sparsely represent images is a challenge for accurately reconstructing magnetic resonance (MR) images from highly undersampled data with noise. In order to further improve the reconstruction quality of the MR image, this paper first proposes tight frame-based group sparsity (TFGS) regularization that can capture the structure information of images appropriately, then TFGS regularization is employed to the image cartoon-texture decomposition model to construct CSMRI algorithm, termed cartoon-texture decomposition CSMRI algorithm (CD-MRI). CD-MRI effectively integrates the total variation and TFGS regularization into the image cartoon-texture decomposition framework, and utilizes the sparse priors of image cartoon and texture components to reconstruct MR images. Virtually, CD-MRI exploits the global sparse representations of image cartoon components by the total variation regularization, and explores group sparse representations of image texture components via the adaptive tight frame learning technique and group sparsity regularization. The alternating iterative method combining with the majorization-minimization algorithm is applied to solve the formulated optimization problem. Finally, simulation experiments demonstrate that the proposed algorithm can achieve high-quality MR image reconstruction from undersampled K-space data, and can remove noise in different sampling schemes. Compared to the previous CSMRI algorithms, the proposed approach can lead to better image reconstruction performance and better robustness to noise.

Automated lesion detectors in retinal fundus images.

Diabetic retinopathy (DR) is a sight-threatening condition occurring in persons with diabetes, which causes progressive damage to the retina. The early detection and diagnosis of DR is vital for saving the vision of diabetic persons. The early signs of DR which appear on the surface of the retina are the dark lesions such as microaneurysms (MAs) and hemorrhages (HEMs), and bright lesions (BLs) such as exudates. In this paper, we propose a novel automated system for the detection and diagnosis of these retinal lesions by processing retinal fundus images. We devise appropriate binary classifiers for these three different types of lesions. Some novel contextual/numerical features are derived, for each lesion type, depending on its inherent properties. This is performed by analysing several wavelet bands (resulting from the isotropic undecimated wavelet transform decomposition of the retinal image green channel) and by using an appropriate combination of Hessian multiscale analysis, variational segmentation and cartoon+texture decomposition. The proposed methodology has been validated on several medical datasets, with a total of 45,770 images, using standard performance measures such as sensitivity and specificity. The individual performance, per frame, of the MA detector is 93% sensitivity and 89% specificity, of the HEM detector is 86% sensitivity and 90% specificity, and of the BL detector is 90% sensitivity and 97% specificity. Regarding the collective performance of these binary detectors, as an automated screening system for DR (meaning that a patient is considered to have DR if it is a positive patient for at least one of the detectors) it achieves an average 95-100% of sensitivity and 70% of specificity at a per patient basis. Furthermore, evaluation conducted on publicly available datasets, for comparison with other existing techniques, shows the promising potential of the proposed detectors.