Noah Lee

A Partial Intensity Invariant Feature Descriptor for Multimodal Retinal Image Registration

Detection of vascular bifurcations is a challenging task in multimodal retinal image registration. Existing algorithms based on bifurcations usually fail in correctly aligning poor quality retinal image pairs. To solve this problem, we propose a novel highly distinctive local feature descriptor named partial intensity invariant feature descriptor (PIIFD) and describe a robust automatic retinal image registration framework named Harris-PIIFD. PIIFD is invariant to image rotation, partially invariant to image intensity, affine transformation, and viewpoint/perspective change. Our Harris-PIIFD framework consists of four steps. First, corner points are used as control point candidates instead of bifurcations since corner points are sufficient and uniformly distributed across the image domain. Second, PIIFDs are extracted for all corner points, and a bilateral matching technique is applied to identify corresponding PIIFDs matches between image pairs. Third, incorrect matches are removed and inaccurate matches are refined. Finally, an adaptive transformation is used to register the image pairs. PIIFD is so distinctive that it can be correctly identified even in nonvascular areas. When tested on 168 pairs of multimodal retinal images, the Harris-PIIFD far outperforms existing algorithms in terms of robustness, accuracy, and computational efficiency.

Towards heterogeneous temporal clinical event pattern discovery: a convolutional approach

Large collections of electronic clinical records today provide us with a vast source of information on medical practice. However, the utilization of those data for exploratory analysis to support clinical decisions is still limited. Extracting useful patterns from such data is particularly challenging because it is longitudinal, sparse and heterogeneous. In this paper, we propose a Nonnegative Matrix Factorization (NMF) based framework using a convolutional approach for open-ended temporal pattern discovery over large collections of clinical records. We call the method One-Sided Convolutional NMF (OSC-NMF). Our framework can mine common as well as individual shift-invariant temporal patterns from heterogeneous events over different patient groups, and handle sparsity as well as scalability problems well. Furthermore, we use an event matrix based representation that can encode quantitatively all key temporal concepts including order, concurrency and synchronicity. We derive efficient multiplicative update rules for OSC-NMF, and also prove theoretically its convergence. Finally, the experimental results on both synthetic and real world electronic patient data are presented to demonstrate the effectiveness of the proposed method.

A Framework for Mining Signatures from Event Sequences and Its Applications in Healthcare Data

This paper proposes a novel temporal knowledge representation and learning framework to perform large-scale temporal signature mining of longitudinal heterogeneous event data. The framework enables the representation, extraction, and mining of high-order latent event structure and relationships within single and multiple event sequences. The proposed knowledge representation maps the heterogeneous event sequences to a geometric image by encoding events as a structured spatial-temporal shape process. We present a doubly constrained convolutional sparse coding framework that learns interpretable and shift-invariant latent temporal event signatures. We show how to cope with the sparsity in the data as well as in the latent factor model by inducing a double sparsity constraint on the β-divergence to learn an overcomplete sparse latent factor model. A novel stochastic optimization scheme performs large-scale incremental learning of group-specific temporal event signatures. We validate the framework on synthetic data and on an electronic health record dataset.

Lipofuscin and autofluorescence metrics in progressive STGD.

PURPOSE To evaluate Stargardt disease (STGD) progression and relative lipofuscin levels via autofluorescence image analysis. METHODS The relationship between focally increased autofluorescence (FIAF), geographic atrophy (GA) and focally decreased autofluorescence (FDAF) was analyzed in serial, registered autofluorescence (AF) scans of 10 patients with STGD (20 eyes, 40 scans; mean follow-up, 2.0 years) using automated techniques. RESULTS GA progressed uniformly in a transition zone with minimal FIAF. Only 4.3% of FIAF progressed to GA or FDAF, despite significant progression of GA (median 30%/year) and FDAF (mean, 29%/year). As a spatial predictor, the mean chance of FIAF for progression to FDAF was 4.3% +/- 4.4%, significantly less than that of random areas (6.7% +/- 4.0%, P = 0.029, Mann-Whitney test). In the seven eyes with GA, the mean chance of FIAF in the transition zone for transition to GA was 12% +/- 8.9%, significantly less than that of random areas (33% +/- 3.6%, P = 0.026, Mann-Whitney test). CONCLUSIONS Autofluorescent flecks and FIAF deposits with AF levels elevated above the initial macular background were less likely in the short term (2 years) to transform to GA and FDAF (AF levels below the final background) than random areas, suggesting additional mechanisms beyond direct lipofuscin toxicity. FIAF/FDAF levels were observed to fluctuate, with focal remodeling of FIAF and FDAF, or rarely, even transition of FDAF to FIAF. FDAF tended to develop, not coincident with, but adjacent to initial FIAF. Because AF identifies these characteristic biological markers so specifically, autofluorescence metrics merit consideration in the study of STGD.

Recovery of macular pigment spectrum in vivo using hyperspectral image analysis

We investigated the feasibility of a novel method for hyperspectral mapping of macular pigment (MP) in vivo. Six healthy subjects were recruited for noninvasive imaging using a snapshot hyperspectral system. The three-dimensional full spatial-spectral data cube was analyzed using non-negative matrix factorization (NMF), wherein the data was decomposed to give spectral signatures and spatial distribution, in search for the MP absorbance spectrum. The NMF was initialized with the in vitro MP spectrum and rank 4 spectral signature decomposition was used to recover the MP spectrum and optical density in vivo. The recovered MP spectra showed two peaks in the blue spectrum, characteristic of MP, giving a detailed in vivo demonstration of these absorbance peaks. The peak MP optical densities ranged from 0.08 to 0.22 (mean 0.15+∕-0.05) and became spatially negligible at diameters 1100 to 1760 μm (4 to 6 deg) in the normal subjects. This objective method was able to exploit prior knowledge (the in vitro MP spectrum) in order to extract an accurate in vivo spectral analysis and full MP spatial profile, while separating the MP spectra from other ocular absorbers. Snapshot hyperspectral imaging in combination with advanced mathematical analysis provides a simple cost-effective approach for MP mapping in vivo.

Dynamic soft drusen remodelling in age-related macular degeneration.

Aims To demonstrate and quantify the dynamic remodelling process of soft drusen resorption and new drusen formation in age-related macular degeneration (AMD) with novel interactive methods. Methods Twenty patients with large soft drusen bilaterally and without advanced AMD were imaged at baseline and again at a mean interval of 2 years (40 eyes, 80 images). Each of the 40 serial pairs of images was precisely registered by an automated technique. The drusen were segmented by a user-interactive method based on a background levelling algorithm and classified into three groups: new drusen (only in the final image), resorbed drusen (present initially but not in the final image) and stable drusen (present in both images). We measured each of these classes as well as the absolute change in drusen |D1 − D0| and the dynamic drusen activity (creation and resorption) Dnew+Dresorbed. Results Mean dynamic activity for the right eye (OD) was 7.33±5.50%, significantly greater than mean absolute change (2.71±2.89%, p=0.0002, t test), with similar results for the left eye (OS). However, dynamic activity OD compared with OS (mean 7.33±5.50 vs 7.91±4.16%, NS) and absolute net change OD versus OS (2.71±2.89 vs 3.46±3.97%, NS) tended to be symmetrical between fellow eyes. Conclusions Dynamic remodelling processes of drusen resorption and new drusen formation are distinct disease activities that can occur simultaneously and are not captured by change in total drusen load. Dynamic changes occur at rates more than twice that of net changes, and may be a useful marker of disease activity.

Learning non-homogenous textures and the unlearning problem with application to drusen detection in retinal images

In this work we present a novel approach for learning non- homogenous textures without facing the unlearning problem. Our learning method mimics the human behavior of selective learning in the sense of fast memory renewal. We perform probabilistic boosting and structural similarity clustering for fast selective learning in a large knowledge domain acquired over different time steps. Applied to non- homogenous texture discrimination, our learning method is the first approach that deals with the unlearning problem applied to the task of drusen segmentation in retinal imagery, which itself is a challenging problem due to high variability of non-homogenous texture appearance. We present preliminary results.

Mining Electronic Medical Records to Explore the Linkage between Healthcare Resource Utilization and Disease Severity in Diabetic Patients

Knowledge discovery in electronic health records (EHRs) is a central aspect for improved clinical decision making, prognosis, and patient management. While EHRs show great promise towards better data integration, automated access, and clinical workflow improvement, the vast information they capture over time pose challenges not only for medical practitioners, but also for the information analysis by machines. The objective of this paper is to promote and emphasize the importance of exploratory analytics that are commensurate with human capabilities and constraints. Within this realm we present a novel temporal event matrix representation and learning framework that discovers complex latent event patterns, which are easily interpretable by humans. We demonstrate our framework on synthetic data and on EHRs together with an extensive validation involving over 70,000 computed latent factor models. The present study is the first to link temporal patterns of healthcare resource utilization (HRU) against a diabetic disease complications severity index to better understand the relationships between disease severity and care delivery.

A hybrid segmentation approach for geographic atrophy in fundus auto-fluorescence images for diagnosis of age-related macular degeneration

Fundus auto-fluorescence (FAF) images with hypo-fluorescence indicate geographic atrophy (GA) of the retinal pigment epithelium (RPE) in age-related macular degeneration (AMD). Manual quantification of GA is time consuming and prone to inter- and intra-observer variability. Automatic quantification is important for determining disease progression and facilitating clinical diagnosis of AMD. In this paper we describe a hybrid segmentation method for GA quantification by identifying hypo-fluorescent GA regions from other interfering retinal vessel structures. First, we employ background illumination correction exploiting a non-linear adaptive smoothing operator. Then, we use the level set framework to perform segmentation of hypo-fluorescent areas. Finally, we present an energy function combining morphological scale-space analysis with a geometric model-based approach to perform segmentation refinement of false positive hypo- fluorescent areas due to interfering retinal structures. The clinically apparent areas of hypo-fluorescence were drawn by an expert grader and compared on a pixel by pixel basis to our segmentation results. The mean sensitivity and specificity of the ROC analysis were 0.89 and 0.98%.

Interactive segmentation for geographic atrophy in retinal fundus images

Fundus auto-fluorescence (FAF) imaging is a non-invasive technique for in vivo ophthalmoscopic inspection of age-related macular degeneration (AMD), the most common cause of blindness in developed countries. Geographic atrophy (GA) is an advanced form of AMD and accounts for 12-21% of severe visual loss in this disorder. Automatic quantification of GA is important for determining disease progression and facilitating clinical diagnosis of AMD. The problem of automatic segmentation of pathological images still remains an unsolved problem. In this paper we leverage the watershed transform and generalized non-linear gradient operators for interactive segmentation and present an intuitive and simple approach for geographic atrophy segmentation. We compare our approach with the state of the art random walker algorithm for interactive segmentation using ROC statistics. Quantitative evaluation experiments on 100 FAF images show a mean sensitivity / specificity of 98.3 / 97.7% for our approach and a mean sensitivity / specificity of 88.2 / 96.6% for the random walker algorithm.