Abstract

Annotations are critical for machine learning and developing Computer Aided Detection (CAD) algorithms. However, a majority of medical data is either unlabeled or annotated only at the image-level. This poses a problem specifically for employing deep learning based approaches for CAD development as they require large amounts of annotated data for training. Data augmentation is a popular solution to address this need. We explore crowdsourcing as a solution for training a deep neural network (DNN) for lesion detection. Our solution employs a strategy to overcome the noisy nature of crowdsourced annotations by i) assigning a reliability factor for each subject of the crowd based on their performance (at global and local levels) and experience and ii) requiring region of interest(ROI) markings rather than pixel-level markings from the crowd. We present a solution for training the DNN with data drawn from a heterogeneous mixture of annotations, namely,very limited number of pixel-level markings by experts and crowdsourced ROI markings. Experimental results obtained for hard exudate detection from color fundus images show that training with processed/refined crowdsourced data is effective as detection performance improves by 25% over training with just expert-markings and by 11% over training with annotation derived using majority voting among the crowd

Abstract

Segmentation of retinal layers from Optical Coherence Tomography (OCT) volumes is a fundamental problem for any computer aided diagnostic algorithm development.This requires preprocessing steps such as denoising, region of interest extraction, flattening and edge detection all of which involve separate parameter tuning. In this paper,we explore deep learning techniques to automate all these steps and handle the presence/absence of pathologies. A model is proposed consisting of a combination of Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM). The CNN is used to extract layers of interest image and extract the edges, while the LSTM is used to trace the layer boundary. This model is trained on a mixture of normal and AMD cases using minimal data. Validation results on three public datasets show that the pixel-wise mean absolute error obtained with our system is1.30±0.48whichis lower than the inter-marker error of1.79±0.76. Our model’s performance is also on par with the existing methods

Abstract

Background and objective: Accurate segmentation of optic disc and cup from monocular color fundus images plays a significant role in the screening and diagnosis of glaucoma. Though optic cup is characterized by the drop in depth from the disc boundary, most existing methods segment the two structures separately and rely only on color and vessel kink based cues due to the lack of explicit depth information in color fundus images. Methods: We propose a novel boundary-based Conditional Random Field formulation that extracts both the optic disc and cup boundaries in a single optimization step. In addition to the color gradients, the proposed method explicitly models the depth which is estimated from the fundus image itself using a coupled, sparse dictionary trained on a set of image-depth map (derived from Optical Coherence Tomography) pairs. Results: The estimated depth achieved a correlation coefficient of 0.80 with respect to the ground truth. The proposed segmentation method outperformed several state-of-the-art methods on five public datasets. The average dice coefficient was in the range of 0.87–0.97 for disc segmentation across three datasets and 0.83 for cup segmentation on the DRISHTI-GS1 test set. The method achieved a good glaucoma classification performance with an average AUC of 0.85 for five fold cross-validation on RIM-ONE v2.

Abstract

Intra-retinal layer segmentation of Optical Coherence Tomography images is critical in the assessment of ocular diseases. Existing Energy minimization based methods employ handcrafted cost terms to define their energy and are not robust to the presence of abnormalities.We propose a novel, Linearly Parameterized, Conditional Random Field(LP-CRF) model whose energy is learnt from a set of training images in an end-to-end manner. The proposed LP-CRF comprises two convolution filter banks to capture the appearance of each tissue region and boundary, the relative weights of the shape priors and an additional term based on the appearance similarity of the adjacent boundary points. All the energy terms are jointly learnt using the Structured Support Vector Machine. The proposed method segments all retinal boundaries in a single step. Our method was evaluated on 107 Normal and 220 AMDB-scan images and found to outperform three publicly available OCT segmentation software. The average unsigned boundary localization error is 1.52±0.29 pixels for segmentation of 8 boundaries on the Normal data set and 1.9±0.65 pixels for 3 boundaries on the combined AMD and Normal data set establishing the robustness of the proposed method.

Abstract

Computer-assisted diagnostic (CAD) tools are of interest as they enable efficient decision-making in clinics and the screening of diseases. The traditional approach to CAD algorithm design focuses on the automated detection of abnormalities independent of the end-user, who can be an image reader or an expert.We propose a reader-centric system design wherein a reader’s attention is drawn to abnormal regions in a least-obtrusive yet effective manner, using saliency-based emphasis of abnormalities and without altering the appearance of the background tissues. We present an assistive lesion-emphasis system (ALES) based on the above idea, for fundus image-based diabetic retinopathy diagnosis. Lesion-saliency is learnt using a convolutional neural network (CNN), inspired by the saliency model of Itti and Koch. The CNN is used to fine-tune standard low-level filters and learn high-level filters for deriving a lesion-saliency map, which is then used to perform lesion-emphasis via a spatially variant version of gamma correction. The proposed system has been evaluated on public datasets and bench marked against other saliency models. It was found to outperform other saliency models by 6% to 30% and boost the contrast-to-noise ratio of lesions by more than 30%. Results of a perceptual study also underscore the effectiveness and, hence, the potential of ALES as an assistive tool for readers

Abstract

In this paper, we propose an end-to-end trainable Convolutional Neural Network (CNN) architecture called the M-net,for segmenting deep (human) brain structures from Magnetic Resonance Images (MRI). A novel scheme is used to learn to combine and represent 3D context information of a given slice in a 2D slice. Consequently, the M-netutilizes only 2D convolution though it operates on 3D data, which makes M-net memory efficient. The segmentation method is evaluated on two publicly available datasets and is compared against publicly available model based segmentation algorithms as well as other classification based algorithms such as Random For-rest and 2D CNN based approaches. Experiment results show that the M-net out performs all these methods in terms of dice coefficient and is at least 3 times faster than other methods in segmenting a new volume which is attractive for clinical use

Abstract

Retinal Neovascularization (NV) is a critical stage of Diabetic Retinopathy (DR) and its detection is important to prevent blindness. Existing fully supervised frameworks typically take a patch-based approach and report good results only on limited number of images due to sparsity of annotated data.We propose a patch-based semi-supervised framework which paves the way for including unlabeled data in training. In this framework, NV patches are modeled using oriented energy and vesselness based features. These features are fused within a co-training based semi-supervised framework by using neighborhood information in feature space. Rule-based criteria on patch-level neovascularity scores is used to derive the final image-level decision. The proposed approach was evaluated on 1 private and 3 public datasets, both at patch and image level detection on nearly 200,000 patches. An AUCof 0.985 with sensitivity of 96.2% at specificity of 92.6% was obtained for abnormality detection at patch-level, while at the image-level, a sensitivity of 96.76% at a specificity of 91.85%were obtained. The achieved performance on a large number of patches indicates the robustness of our approach.

Abstract

Retinal vessel keypoint detection and classification is a fundamental step in tracking the physiological changes that occur in the retina which is linked to various retinal and systemic diseases. In this paper, we propose a novel Vessel Keypoint Detector (VKD) which is derived from the projection of log-polar transformed binary patches around vessel points.VKD is used to design a two stage solution for junction detection and classification. In the first stage, the keypoints detected using VKD are refined using curvature orientation information to extract candidate junctions. True junctions from these candidates are identified in a supervised manner using a Random Forest classifier. In the next stage, a novel combination of local orientation and shape based features is extracted from the junction points and classified using asecond Random Forest classifier. Evaluation results on five datasets show that the designed system is robust to changes in resolution and other variations across datasets, with average values of accuracy/sensitivity/specificity for junction detection being 0.78/0.79/0.75 and for junction classification being 0.87/0.85/0.88. Our system outperforms the state of the art method [1] by at least 11%, on the DRIVE and IOSTAR datasets. These results demonstrate the effectiveness of VKDfor vessel analysis.

Abstract

Automated segmentation of cortical and noncortical human brain structures has been hither to approached using nonrigid registration followed by label fusion. We propose an alternative approach for this using a convolutional neural network (CNN) which classifies a voxel into one of many structures. Four different kinds of two-dimensional and three-dimensional intensity patches are extracted for each voxel, providing local and global (context) information to the CNN. The proposed approach is evaluated on five different publicly available datasets which differ in the number of labels per volume. The obtained mean Dice coefficient varied according to the number of labels, for example, it is0.844
0.031and0.743
0.019for datasets with the least(32) and the most (134) number of labels, respectively. These figures are marginally better or on par with those obtained with the current state-of-the-art methods on nearly all datasets, at a reduced computational time. The consistently good performance of the proposed method across datasets and no requirement for registration make it attractive for many applications where reduced computational time is necessary.

Abstract

Saliency computation is widely studied in computer vision but not in medical imaging. Existing computational saliency models have been developed for general (natural) images and hence may not be suitable for medical images. This is due to the variety of imaging modalities and the requirement of the models to capture not only normal but also deviations from normal anatomy. We present a biologically inspired model for colour fundus images and illustrate it for the case of diabetic retinopathy. The proposed model uses spatially-varying morphological operations to enhance lesions locally and combines an ensemble of results, of such operations, to generate the saliency map. The model is validated against an average Human Gaze map of 15 experts and found to have 10% higher recall (at 100% precision) than four leading saliency models proposed for natural images. The F-score for match with manual lesion markings by 5 experts was0.4 (as opposed to 0.532 for gaze map) for our model and very poor for existing models. The model’s utility is shown via a novel enhancement method which employs saliency to selectively enhance the abnormal regions and this was found to boost their contrast to noise ratio by∼30%