Joo Hwee Lim

Automatic Optic Disc Detection From Retinal Images by a Line Operator

Under the framework of computer-aided eye disease diagnosis, this paper presents an automatic optic disc (OD) detection technique. The proposed technique makes use of the unique circular brightness structure associated with the OD, i.e., the OD usually has a circular shape and is brighter than the surrounding pixels whose intensity becomes darker gradually with their distances from the OD center. A line operator is designed to capture such circular brightness structure, which evaluates the image brightness variation along multiple line segments of specific orientations that pass through each retinal image pixel. The orientation of the line segment with the minimum/maximum variation has specific pattern that can be used to locate the OD accurately. The proposed technique has been tested over four public datasets that include 130, 89, 40, and 81 images of healthy and pathological retinas, respectively. Experiments show that the designed line operator is tolerant to different types of retinal lesion and imaging artifacts, and an average OD detection accuracy of 97.4% is obtained.

Automated Layer Segmentation of Optical Coherence Tomography Images

By measuring the thickness of the retinal nerve fiber layer, retinal optical coherence tomography (OCT) images are now increasingly used for the diagnosis of glaucoma. This paper reports an automatic OCT layer segmentation technique that can be used for computer-aided glaucoma diagnosis. In the proposed technique, blood vessels are first detected through an iterative polynomial smoothing procedure. OCT images are then filtered by a bilateral filter and a median filter sequentially. In particular, both filters suppress the local image noise but the bilateral filter has a special characteristic that keeps the global trend of the image value variation. After the image filtering, edges are detected and the edge segments corresponding to the layer boundary are further identified and clustered to form the layer boundary. Experiments over OCT images of four subjects show that the proposed technique segments layers of OCT images efficiently.

Similarity Learning for Nearest Neighbor Classification

In this paper, we propose an algorithm for learning a general class of similarity measures for kNN classification. This class encompasses, among others, the standard cosine measure, as well as the Dice and Jaccard coefficients. The algorithm we propose is an extension of the voted perceptron algorithm and allows one to learn different types of similarity functions (either based on diagonal, symmetric or asymmetric similarity matrices). The results we obtained show that learning similarity measures yields significant improvements on several collections, for two prediction rules: the standard kNN rule, which was our primary goal, and a symmetric version of it.

Optic disc region of interest localization in fundus image for Glaucoma detection in ARGALI

Glaucoma is the second leading cause of permanent blindness worldwide. Glaucoma can be diagnosed through measurement of neuro-retinal optic cup-to-disc ratio (CDR). Correctly determining the optic disc region of interest (ROI) will produce a smaller initial image which takes much lesser time taken to process compared to the entire image. The earlier ROI localization in the ARGALI system used a grid based method. The new algorithm adds a preprocessing step before analyzing the image. This step significantly improves the performance of the ROI detection. A batch of 1564 retinal images from the Singapore Eye Research Centre was used to compare the performance of the two methods. From the results, the earlier and new algorithm detects the ROI correctly for 88% and 96% of the images respectively. The results indicate potential applicability of the method for automated and objective mass screening for early detection of glaucoma.

A Computer-Aided Diagnosis System of Nuclear Cataract

Cataracts are the leading cause of blindness worldwide, and nuclear cataract is the most common form of cataract. An algorithm for automatic diagnosis of nuclear cataract is investigated in this paper. Nuclear cataract is graded according to the severity of opacity using slit lamp lens images. Anatomical structure in the lens image is detected using a modified active shape model. On the basis of the anatomical landmark, local features are extracted according to clinical grading protocol. Support vector machine regression is employed for grade prediction. This is the first time that the nucleus region can be detected automatically in slit lamp images. The system is validated using clinical images and clinical ground truth on >5000 images. The success rate of structure detection is 95% and the average grading difference is 0.36 on a 5.0 scale. The automatic diagnosis system can improve the grading objectivity and potentially be used in clinics and population studies to save the workload of ophthalmologists.

Convex hull based neuro-retinal optic cup ellipse optimization in glaucoma diagnosis

Glaucoma is the second leading cause of blindness. Glaucoma can be diagnosed through measurement of neuro-retinal optic cup-to-disc ratio (CDR). Automatic calculation of optic cup boundary is challenging due to the interweavement of blood vessels with the surrounding tissues around the cup. A Convex Hull based Neuro-Retinal Optic Cup Ellipse Optimization algorithm improves the accuracy of the boundary estimation. The algorithm’s effectiveness is demonstrated on 70 clinical patient’s data set collected from Singapore Eye Research Institute. The root mean squared error of the new algorithm is 43% better than the ARGALI system which is the state-of-the-art. This further leads to a large clinical evaluation of the algorithm involving 15 thousand patients from Australia and Singapore.

An ensemble classifier learning approach to ROC optimization

An ensemble learning framework is proposed to optimize the receiver operating characteristic (ROC) curve corresponding to a given classifier. The proposed ensemble maximal figure-of-merit (E-MFoM) learning framework meets four key requirements desirable for ROC optimization, namely: (1) each classifier in the ensemble can be learned with any specified performance metric for any given classifier design; (2) such a classifier is discriminative in nature and attempts to optimize a particular operating point on the ROC curve of the classifier; (3) an ensemble approximation to the overall behavior of the ROC curve can be established by sampling a set of operating points; and (4) ensemble decision rules can be formulated by grouping these sampled classifiers with a uniform scoring function. We evaluate the proposed framework using 3 testing databases, the Reuters and two UCI sets. Our experimental results clearly show that E-MFoM learning outperforms the state-of-the-art algorithms using Wilcoxon-Mann-Whitney rank statistics

Multi-level local feature classification for bleeding detection in Wireless Capsule Endoscopy images

This paper presents a novel multi-level approach for bleeding detection in Wireless Capsule Endoscopy (WCE) images. In the low-level processing, each cell of K×K pixels is characterized by an adaptive color histogram which optimizes the information representation for WCE images. A Neural Network (NN) cell-classifier is trained to classify cells in an image as bleeding or non-bleeding patches. In the intermediate-level processing, a block which covers 3×3 cells is formed. The intermediate-level representation of the block is generated from the low-level classifications of the cells, which captures the spatial local correlations of the cell classifications. Again, a NN block-classifier is trained to classify the blocks as bleeding or non-bleeding ones. In the high-level processing, the low-level cell-based and intermediate-level block-based classifications are fused for final detection. In this way, our approach can combine the low-level features from pixels and intermediate-level features from local regions to achieve robust bleeding detection. Experiments on real WCE videos have shown that the proposed method of multi-level classification is not only accurate in both detection and localization of potential bleedings in WCE images but also robust to complex local noisy features.

Character Recognition in Natural Scenes Using Convolutional Co-occurrence HOG

Recognition of characters in natural images is a challenging task due to the complex background, variations of text size and perspective distortion, etc. Traditional optical character recognition (OCR) engine cannot perform well on those unconstrained text images. A novel technique is proposed in this paper that makes use of convolutional cooccurrence histogram of oriented gradient (ConvCoHOG), which is more robust and discriminative than both the histogram of oriented gradient (HOG) and the co-occurrence histogram of oriented gradients (CoHOG). In the proposed technique, a more informative feature is constructed by exhaustively extracting features from every possible image patches within character images. Experiments on two public datasets including the ICDAr 2003 Robust Reading character dataset and the Street View Text (SVT) dataset, show that our proposed character recognition technique obtains superior performance compared with state-of-the-art techniques.

Epitomized summarization of wireless capsule endoscopic videos for efficient visualization

A video recording of an examination by Wireless Capsule Endoscopy (WCE) may typically contain more than 55,000 video frames, which makes the manual visual screening by an experienced gastroenterologist a highly time-consuming task. In this paper, we propose a novel method of epitomized summarization of WCE videos for efficient visualization to a gastroenterologist. For each short sequence of a WCE video, an epitomized frame is generated. New constraints are introduced into the epitome formulation to achieve the necessary visual quality for manual examination, and an EM algorithm for learning the epitome is derived. First, the local context weights are introduced to generate the epitomized frame. The epitomized frame preserves the appearance of all the input patches from the frames of the short sequence. Furthermore, by introducing spatial distributions for semantic interpretation of image patches in our epitome formulation, we show that it also provides a framework to facilitate the semantic description of visual features to generate organized visual summarization of WCE video, where the patches in different positions correspond to different semantic information. Our experiments on real WCE videos show that, using epitomized summarization, the number of frames have to be examined by the gastroenterologist can be reduced to less than one-tenth of the original frames in the video.