David Dagan Feng

University of Sydney, Sydney, New South Wales, Australia

Are you David Dagan Feng?

Claim your profile

Publications (213)131.26 Total impact

  • [Show abstract] [Hide abstract]
    ABSTRACT: In this paper, we propose a new Locality-constrained Subcluster Representation Ensemble (LSRE) model, to classify high-resolution computed tomography (HRCT) images of interstitial lung diseases (ILDs). Medical images normally exhibit large intra-class variation and inter-class ambiguity in the feature space. Modelling of feature space separation between different classes is thus problematic and this affects the classification performance. Our LSRE model tackles this issue in an ensemble classification construct. The image set is first partitioned into subclusters based on spectral clustering with approximation-based affinity matrix. Basis representations of the test image are then generated with sparse approximation from the subclusters. These basis representations are finally fused with approximation- and distribution-based weights to classify the test image. Our experimental results on a large HRCT database show good performance improvement over existing popular classifiers. Copyright © 2015 Elsevier B.V. All rights reserved.
    Medical image analysis 03/2015; 22(1). DOI:10.1016/j.media.2015.03.003 · 3.68 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Content-based image retrieval (CBIR) has been applied to a variety of medical applications, e.g., pathology research and clinical decision support, and bag-of-features (BOF) model is one of the most widely used techniques. In this study, we address the problem of vocabulary pruning to reduce the influence from the redundant and noisy visual words. The conditional probability of each word upon the hidden topics extracted using probabilistic Latent Semantic Analysis (pLSA) is firstly calculated. A ranking method is then proposed to compute the significance of the words based on the relationship between the words and topics. Experiments on the publicly available Early Lung Cancer Action Program (ELCAP) database show that the method can reduce the number of words required while improving the retrieval performance. The proposed method is applicable to general image retrieval since it is independent of the problem domain.
    Artificial Life and Computational Intelligence, 02/2015: pages 436-445;
  • [Show abstract] [Hide abstract]
    ABSTRACT: The rapid growth of video data demands both effective and efficient video summarization methods so that users are empowered to quickly browse and comprehend a large amount of video content. In this paper, we formulate the video summarization task with a novel minimum sparse reconstruction (MSR) problem. That is, the original video sequence can be best reconstructed with as few selected keyframes as possible. Different from the recently proposed convex relaxation based sparse dictionary selection method, our proposed method utilizes the true sparse constraint L-0 norm, instead of the relaxed constraint L-2,L-1 norm, such that keyframes are directly selected as a sparse dictionary that can well reconstruct all the video frames. An on-line version is further developed owing to the real-time efficiency of the proposed MSR principle. In addition, a percentage of reconstruction (POR) criterion is proposed to intuitively guide users in obtaining a summary with an appropriate length. Experimental results on two benchmark datasets with various types of videos demonstrate that the proposed methods outperform the state of the art.
    Pattern Recognition 02/2015; 48(2). DOI:10.1016/j.patcog.2014.08.002 · 2.58 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Medical images usually exhibit large intra-class variation and inter-class ambiguity in the feature space, which could affect classification accuracy. To tackle this issue, we propose a new Large Margin Local Estimate (LMLE) classification model with sub-categorization based sparse representation. We first sub-categorize the reference sets of different classes into multiple clusters, to reduce feature variation within each subcategory compared to the entire reference set. Local estimates are generated for the test image using sparse representation with reference subcategories as the dictionaries. The similarity between the test image and each class is then computed by fusing the distances with the local estimates in a learning-based large margin aggregation construct to alleviate the problem of inter-class ambiguity. The derived similarities are finally used to determine the class label. We demonstrate that our LMLE model is generally applicable to different imaging modalities, and applied it to three tasks: interstitial lung disease (ILD) classification on high-resolution computed tomography (HRCT) images, phenotype binary classification and continuous regression on brain magnetic resonance (MR) imaging. Our experimental results show statistically significant performance improvements over existing popular classifiers.
    IEEE Transactions on Medical Imaging 01/2015; 34(6). DOI:10.1109/TMI.2015.2393954 · 3.80 Impact Factor
  • Lin Shu, Xiaoming Tao, David Dagan Feng
    [Show abstract] [Hide abstract]
    ABSTRACT: Resistive sensor arrays have been increasingly adopted in wearable electronic applications, which require low-complexity and low-energy circuits. However, current readout strategies for resistive sensor arrays require additional electrical components, such as transistors, diodes, multiplexers, op-amps, switches, current sources, and A/D converters, leading to a considerable increase in circuit complexity, power consumption, system instability, and crosstalk error. To address the problem, this paper proposes a new approach, which determines sensor resistance values by establishing and solving resistance matrix equations of sensor arrays. Unlike conventional approaches, it allows crosstalk currents in arrays to avoid additional components that are originally used for eliminating crosstalk currents and minimizing crosstalk error. Meanwhile, it takes advantage of on-chip resources of wearable platforms, thereby reducing redundant chips. It was implemented on a prototype of 10 × 10 textile resistive sensor array, which was taken in a sensing cushion for sitting pressure monitoring of chair bound people. Experimental results on this array platform showed the new approach achieved a satisfactory accuracy (0.61% ± 0.41%), as well as a low crosstalk error (2.77% ± 0.61%). The fabricated sensing cushion also exhibited a relatively low pressure measurement error (6.30% ± 0.75%). Compared with other approaches, the proposed approach demonstrated the lowest circuit complexity on a microcontroller based wearable platform, and a sufficient sensor capacity. It is ideal for a wide range of applications like wearable or implantable sensing, presenting a reference for the design of low-complexity and low-crosstalk error wearable systems based on resistive sensor arrays.
    IEEE Sensors Journal 01/2015; 15(1):442-452. DOI:10.1109/JSEN.2014.2333518 · 1.85 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Accurate approximation of noise in hyperspectral (HS) images plays an important role in better visualization and image processing. Conventional algorithms often hypothesize the noise type to be either purely additive or of a mixed noise type for the signal-dependent (SD) noise component and the signal-independent (SI) noise component in HS images. This can result in application-driven algorithm design and limited use in different noise types. Moreover, as the highly textured HS images have abundant edges and textures, existing algorithms may fail to produce accurate noise estimation. To address these challenges, we propose a noise estimation algorithm that can adaptively estimate both purely additive noise and mixed noise in HS images with various complexities. First, homogeneous areas are automatically detected using a new region-growing-based approach, in which the similarity of two pixels is calculated by a robust spectral metric. Then, the mixed noise variance of each homogeneous region is estimated based on multiple linear regression technology. Finally, intensities of the SD and SI noise are obtained with a modified scatter plot approach. We quantitatively evaluated our algorithm on the synthetic HS data. Compared with the benchmarking and state-of-the-art algorithms, the proposed algorithm is more accurate and robust when facing images with different complexities. Experimental results with real Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) images further demonstrated the superiority of our algorithm.
    Applied Optics 10/2014; 53(30). DOI:10.1364/AO.53.007059 · 1.78 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Neuroimaging has played an important role in non-invasive diagnosis and differentiation of neurodegenerative disorders, such as Alzheimer's disease and Mild Cognitive Impairment. Various types of features have been extracted from the neuroimaging data to characterize the disorders, and these features can be roughly divided into global and local features. Recent studies show a tendency of using the local features in disease characterization, since they are capable of identifying the subtle disease-specific patterns associated with the effects of the disease on human brain. However, problems arise if the neuroimaging database involved multiple disorders or progressive disorders, for disorders of different types or at different progressive stages might exhibit different degenerative patterns. It is difficult for the researchers to reach consensus on what brain regions could best distinguish multiple disorders or multiple progression stages. In this study we proposed a Multi-Channel pattern analysis approach to identify the most discriminative local brain metabolism features for neurodegenerative disorder characterization. We compared our method to the global methods and other pattern analysis methods based on clinical expertise or test statistics. The preliminary results suggested that the proposed Multi-Channel pattern analysis method outperformed other approaches in Alzheimer's disease characterization, and meanwhile provided important insights of underlying pathology of Alzheimer's disease and Mild Cognitive Impairment.
    Computerized Medical Imaging and Graphics 09/2014; 38(6). DOI:10.1016/j.compmedimag.2014.05.003 · 1.50 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Multi-modality positron emission tomography and computed tomography (PET-CT) imaging depicts biological and physiological functions (from PET) within a higher resolution anatomical reference frame (from CT). The need to efficiently assimilate the information from these co-aligned volumes simultaneously has resulted in 3D visualisation methods that depict e.g., slice of interest (SOI) from PET combined with direct volume rendering (DVR) of CT. However because DVR renders the whole volume, regions of interests (ROIs) such as tumours that are embedded within the volume may be occluded from view. Volume clipping is typically used to remove occluding structures by `cutting away' parts of the volume; this involves tedious trail-and-error tweaking of the clipping attempts until a satisfied visualisation is made, thus restricting its application. Hence, we propose a new automated opacity-driven volume clipping method for PET-CT using DVR-SOI visualisation. Our method dynamically calculates the volume clipping depth by considering the opacity information of the CT voxels in front of the PET SOI, thereby ensuring that only the relevant anatomical information from the CT is visualised while not impairing the visibility of the PET SOI. We outline the improvements of our method when compared to conventional 2D and traditional DVR-SOI visualisations.
    36th Annual Conference Proceedings of the IEEE Engineering in Medicine & Biology Society (EMBC 2014), Chicago, IL, USA; 08/2014
  • [Show abstract] [Hide abstract]
    ABSTRACT: With the increasing amount of image data available for cancer staging and diagnosis, it is clear that content-based image retrieval techniques are becoming more important to assist physicians in making diagnoses and tracking disease. Domain-specific feature descriptors have been previously shown to be effective in the retrieval of lung tumors. This work proposes a method to improve the rotation invariance of the hierarchical spatial descriptor, as well as presents a new binary descriptor for the retrieval of lung nodule images. The descriptors were evaluated on the ELCAP public access database, exhibiting good performance overall.
  • ACM Transactions on Multimedia Computing Communications and Applications 08/2014; 11(1):1-21. DOI:10.1145/2632267 · 0.90 Impact Factor
  • Lei Bi, Jinman Kim, David Dagan Feng, Michael Fulham
    [Show abstract] [Hide abstract]
    ABSTRACT: Fluorodeoxyglucose positron emission tomography - computed tomography (FDG PET-CT) is the preferred image modality for lymphoma diagnosis. Sites of disease generally appear as foci of increased FDG uptake. Thresholding methods are often applied to robustly separate these regions. However, its main limitation is that it also includes sites of FDG excretion and physiological FDG uptake regions, which we define as FEPU - sites of FEPU include the bladder, renal, papillae, ureters, brain, heart and brown fat. FEPU can make image interpretation problematic. The ability to identify and label FEPU sites and separate them from abnormal regions is an important process that could improve image interpretation. We propose a new method to automatically separate and label FEPU sites from the thresholded PET images. Our method is based on the selective use of features extracted from data types comprising of PET, CT and PET-CT. Our FEPU classification of 43 clinical lymphoma patient studies revealed higher accuracy when compared to non-selective image features.
    36th Annual Conference Proceedings of the IEEE Engineering in Medicine & Biology Society (EMBC 2014), Chicago, IL, USA; 08/2014
  • [Show abstract] [Hide abstract]
    ABSTRACT: Automated and general medical image segmentation can be challenging because the foreground and the background may have complicated and overlapping density distributions in medical imaging. Conventional region-based level-set algorithms often assume piecewise constant (PC) or piecewise smooth (PS) for segments, which are implausible for general medical image segmentation. Further, low contrast and noise make identification of the boundaries between foreground and background difficult for edge-based level set algorithms. Thus, to address these problems, we suggest a supervised variational level set segmentation model to harness the statistical region energy functional with a weighted probability approximation. Our approach models the region density distributions by using the mixture-of-mixtures Gaussian model to better approximate real intensity distributions and distinguish statistical intensity differences between foreground and background. The region-based statistical model in our algorithm can intuitively provide better performance on noisy images. We constructed a weighted probability map on graphs to incorporate spatial indications from user input with a contextual constraint based on the minimization of contextual graphs energy functional. We measured the performance of our approach on 10 noisy synthetic images and 58 medical datasets with heterogeneous intensities and ill-defined boundaries and compared our technique to the Chan-Vese (CV) region-based level set model, the geodesic active contour model with distance regularization (GACD) and the random walker (RW) model. Our method consistently achieved the highest Dice similarity coefficient (DSC) when compared to the other methods.
    IEEE transactions on bio-medical engineering 07/2014; 62(1). DOI:10.1109/TBME.2014.2344660 · 2.23 Impact Factor
  • Biomedical Signal Processing and Control 07/2014; 12. DOI:10.1016/j.bspc.2014.02.003 · 1.53 Impact Factor
  • Tong Zhang, Yong Xia, David Dagan Feng
    [Show abstract] [Hide abstract]
    ABSTRACT: The hidden Markov random field (HMRF) model has been widely used in image segmentation, as it provides a spatially constrained clustering scheme on two sets of random variables. However, in many HMRF-based segmentation approaches, both the latent class labels and statistical parameters have been estimated by deterministic techniques, which usually lead to local convergence and less accurate segmentation. In this paper, we incorporate the immune inspired clonal selection algorithm (CSA) and Markov chain Monte Carlo (MCMC) method into HMRF model estimation, and thus propose the HMRF–CSA algorithm for brain MR image segmentation. Our algorithm employs a three-step iterative process that consists of MCMC-based class labels estimation, bias field correction and CSA-based statistical parameter estimation. Since both the MCMC and CSA are global optimization techniques, the proposed algorithm has the potential to overcome the drawback of traditional HMRF-based segmentation approaches. We compared our algorithm to the state-of-the-art GA–EM algorithm, deformable cosegmentation algorithm, the segmentation routines in the widely-used statistical parametric mapping (SPM) software package and the FMRIB software library (FSL) on both simulated and clinical brain MR images. Our results show that the proposed HMRF–CSA algorithm is robust to image artifacts and can differentiate major brain structures more accurately than other three algorithms.
    Biomedical Signal Processing and Control 07/2014; 12:10–18. DOI:10.1016/j.bspc.2013.07.010 · 1.53 Impact Factor
  • Tong Zhang, Yong Xia, David Dagan Feng
    [Show abstract] [Hide abstract]
    ABSTRACT: Statistical classification of voxels in brain magnetic resonance (MR) images into major tissue types plays an important role in neuroscience research and clinical practices, in which model estimation is an essential step. Despite their prevalence, traditional techniques, such as the expectation–maximization (EM) algorithm and genetic algorithm (GA), have inherent limitations, and may result in less-accurate classification. In this paper, we introduce the immune-inspired clonal selection algorithm (CSA) to the maximum likelihood estimation of the Gaussian mixture model (GMM), and thus propose the GMM-CSA algorithm for automated voxel classification in brain MR images. This algorithm achieves simultaneous voxel classification and bias field correction in a three-stage iterative process under the CSA framework. At each iteration, a population of admissible model parameters, voxel labels and estimated bias field are updated. To explore the prior anatomical knowledge, we also construct a probabilistic brain atlas for each MR study and incorporate the atlas into the classification process. The GMM-CSA algorithm has been compared to five state-of-the-art brain MR image segmentation approaches on both simulated and clinical data. Our results show that the proposed algorithm is capable of classifying voxels in brain MR images into major tissue types more accurately.
    Neurocomputing 06/2014; 134:122–131. DOI:10.1016/j.neucom.2012.12.081 · 2.01 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Many approaches to facial expression recognition utilize only one type of features at a time. It can be difficult for a single type of features to characterize in a best possible way the variations and complexity of realistic facial expressions. In this paper, we propose a spectral embedding based multi-view dimension reduction method to fuse multiple features for facial expression recognition. Facial expression features extracted from one type of expressions can be assumed to form a manifold embedded in a high dimensional feature space. We construct a neighborhood graph that encodes the structure of the manifold locally. A graph Laplacian matrix is constructed whose spectral decompositions reveal the low dimensional structure of the manifold. In order to obtain discriminative features for classification, we propose to build a neighborhood graph in a supervised manner by utilizing the label information of training data. As a result, multiple features are able to be transformed into a unified low dimensional feature space by combining the Laplacian matrix of each view with the multiview spectral embedding algorithm. A linearization method is utilized to map unseen data to the learned unified subspace. Experiments are conducted on a set of established real-world and benchmark datasets. The experimental results provide a strong support to the effectiveness of the proposed feature fusion framework on realistic facial expressions.
    Neurocomputing 04/2014; 129:136–145. DOI:10.1016/j.neucom.2013.09.046 · 2.01 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Positron emission tomography (PET) images usually suffer from poor signal-to-noise ratio (SNR) due to the high level of noise and low spatial resolution, which adversely affect its performance for lesion detection and quantification. The complementary information present in high-resolution anatomical images from multi-modality imaging systems could potentially be used to improve the ability to detect and/or quantify lesions. However, previous methods that use anatomical priors usually require matched organ/lesion boundaries. In this study, we investigated the use of anatomical information to suppress noise in PET images while preserving both quantitative accuracy and the amplitude of prominent signals that do not have corresponding boundaries on computerized tomography (CT). The proposed approach was realized through a postreconstruction filter based on the nonlocal means (NLM) filter, which reduces noise by computing the weighted average of voxels based on the similarity measurement between patches of voxels within the image. Anatomical knowledge obtained from CT was incorporated to constrain the similarity measurement within a subset of voxels. In contrast to other methods that use anatomical priors, the actual number of neighboring voxels and weights used for smoothing were determined from a robust measurement on PET images within the subset. Thus, the proposed approach can be robust to signal mismatches between PET and CT. A 3-D search scheme was also investigated for the volumetric PET/CT data. The proposed anatomically guided median nonlocal means filter (AMNLM) was first evaluated using a computer phantom and a physical phantom to simulate realistic but challenging situations where small lesions are located in homogeneous regions, which can be detected on PET but not on CT. The proposed method was further assessed with a clinical study of a patient with lung lesions. The performance of the proposed method was compared to Gaussian, edge-preserving bilateral and NLM filters, as well as median nonlocal means (MNLM) filtering without an anatomical prior. The proposed AMNLM method yielded improved lesion contrast and SNR compared with other methods even with imperfect anatomical knowledge, such as missing lesion boundaries and mismatched organ boundaries.
    IEEE Transactions on Medical Imaging 03/2014; 33(3):636-50. DOI:10.1109/TMI.2013.2292881 · 3.80 Impact Factor
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Parametric FDG-PET images offer the potential for automated identification of the different dementia syndromes. However, various existing image features and classifiers have their limitations in characterizing and differentiating the patterns of this disease. We reported a hybrid feature extraction, selection, and classification approach, namely, the GA-MKL algorithm, for separating patients with suspected Alzheimer's disease and frontotemporal dementia from normal controls. In this approach, we extracted three groups of features to describe the average level, spatial variation, and asymmetry of glucose metabolic rates in 116 cortical volumes. An optimal combination of features, that is, capable of classifying dementia cases was identified by a genetic algorithm- (GA-) based method. The condition of each FDG-PET study was predicted by applying the selected features to a multikernel learning (MKL) machine, in which the weighting parameter of each kernel function can be automatically estimated. We compared our approach to two state-of-the-art dementia identification algorithms on a set of 129 clinical cases and improved the performance in separating the dementia types, achieving accuracy of 94.62%. There is a very good agreement between the proposed automated technique and the diagnosis made by clinicians.
    BioMed Research International 02/2014; 2014:421743. DOI:10.1155/2014/421743 · 2.71 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: We present a lesion detection and characterization method for 18F-fluorodeoxyglucose positron emission tomography - computed tomography (FDG PET-CT) images of the thorax in the evaluation of patients with primary non-small cell lung cancer (NSCLC) with regional nodal disease. Lesion detection can be difficult due to low contrast between lesions and normal anatomical structures. Lesion characterization is also challenging due to similar spatial characteristics between the lung tumors and abnormal lymph nodes. To tackle these problems, we propose a context driven approximation (CDA) method. There are two main components of our method. First, a sparse representation technique with region-level contexts was designed for lesion detection. To discriminate low-contrast data with sparse representation, we propose a reference consistency constraint and a spatial consistent constraint. Second, a multi-atlas technique with image-level contexts was designed to represent the spatial characteristics for lesion characterization. To accommodate inter-subject variation in a multi-atlas model, we propose an appearance constraint and a similarity constraint. The CDA method is effective with a simple feature set, and does not require parametric modeling of feature space separation. The experiments on a clinical FDG PET-CT dataset show promising performance improvement over the state-of-the-art.
    IEEE Transactions on Medical Imaging 02/2014; 33(2):408-421. DOI:10.1109/TMI.2013.2285931 · 3.80 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Medical images typically exhibit complex feature space distributions due to high intra-class variation and inter-class ambiguity. Monolithic classification models are often problematic. In this study, we propose a novel Large Margin Local Estimate (LMLE) method for medical image classification. In the first step, the reference images are subcategorized, and local estimates of the test image are computed based on the reference subcategories. In the second step, the local estimates are fused in a large margin model to derive the similarity level between the test image and the reference images, and the test image is classified accordingly. For evaluation, the LMLE method is applied to classify image patches of different interstitial lung disease (ILD) patterns on high-resolution computed tomography (HRCT) images. We demonstrate promising performance improvement over the state-of-the-art.

Publication Stats

867 Citations
131.26 Total Impact Points

Institutions

  • 2001–2015
    • University of Sydney
      • • School of Information Technologies
      • • Biomedical and Multimedia Information Technology Research Group (BMIT)
      Sydney, New South Wales, Australia
  • 2014
    • Yale University
      • Department of Diagnostic Radiology and Pediatric Diagnostic Radiology
      New Haven, Connecticut, United States
    • Budapest University of Technology and Economics
      Budapeŝto, Budapest, Hungary
  • 2009–2014
    • Shanghai Jiao Tong University
      Shanghai, Shanghai Shi, China
    • Westmead Hospital
      • Department of Medical Physics
      Sydney, New South Wales, Australia
  • 2012
    • Royal Prince Alfred Hospital
      • Department of Pet & Nuclear Medicine
      Camperdown, New South Wales, Australia
  • 2007–2012
    • The University of Hong Kong
      Hong Kong, Hong Kong
    • Tsinghua University
      • Department of Biomedical Engineering
      Beijing, Beijing Shi, China
  • 1998–2012
    • The Hong Kong Polytechnic University
      • Department of Electronic and Information Engineering
      Hong Kong, Hong Kong
  • 2005
    • Heilongjiang University
      Charbin, Heilongjiang Sheng, China
  • 2000
    • Northwestern Polytechnical University
      Xi’an, Liaoning, China