Search

We describe the importance of identifying pigment networks in lesions which may be melanomas, and survey methods for identifying pigment networks (PN) in dermoscopic images. We then give details of how machine learning can be used to classify images into three classes: PN Absent, Regular PN and Irregular PN.

Skin lesions are often comprised of various colours. The presence of multiple colours with an irregular distribution can signal malignancy. Among common colours under dermoscopy, blue-grey (blue-white veil) is a strong indicator of malignant melanoma. Since it is not always easy to visually identify and recognize this feature, a computerised automatic colour analysis method can provide the clinician with an objective second opinion. In this paper, we put forward an innovative method, through colour analysis and computer vision techniques, to automatically detect and segment blue-white veil areas in dermoscopy images. The proposed method is an attempt to mimic the human perception of lesion colours, and improves and outperforms the state-of-the-art as shown in our experiments.

In this paper we propose a new log-chromaticity 2-D colour space, an extension of previous approaches, which succeeds in removing confounding factors from dermoscopic images: (i) the effects of the particular camera characteristics for the camera system used in forming RGB images; (ii) the colour of the light used in the dermoscope; (iii) shading induced by imaging non-flat skin surfaces; (iv) and light intensity, removing the effect of light-intensity falloff toward the edges of the dermoscopic image. In the context of a blind source separation of the underlying colour, we arrive at intrinsic melanin and hemoglobin images, whose properties are then used in supervised learning to achieve excellent malignant vs. benign skin lesion classification. In addition, we propose using the geometric-mean of colour for skin lesion segmentation based on simple grey-level thresholding, with results outperforming the state of the art.

There is an increasing demand for automated detection and analysis of dermoscopy structures and malignancy clues such as streaks in dermoscopy images, for computer-aided early diagnosis of deadly melanoma. This paper presents a novel approach for streak detection and visualization on dermoscopic images. We tackle the detection of streaks by means of ridge and valley estimation. Orientation estimation and correction is applied to detect low contrast and fuzzy streaks lines, and candidate streaks are used to classify dermoscopy images into streaks Absent or Present with the AUC of 90.5% on 300 dermoscopy images. Our approach can also detect starburst pattern of regular streaks using detected linear structures with accuracy of 81.5% and AUC of 87.7%.

This paper presents a novel approach to analysis and classification of skin lesions based on their growth pattern. Our method constructs a tree structure for every lesion by repeatedly subdividing the image into sub-images using color based clustering. In this method, segmentation which is a challenging task is not required. The obtained multi-scale tree structure provides a framework that allows us to extract a variety of features, based on the appearance of the tree structure or sub-images corresponding to nodes of the tree. Preliminary features (the number of nodes, leaves, and depth of the tree, and 9 compactness indices of the dark spots represented by the sub-images associated with each node of the tree) are used to train a supervised learning algorithm. Results show the strength of the method in classifying lesions into malignant and benign classes. We achieved Precision of 0.855, Recall of 0.849, and F-measure of 0.834 using 3-layer perceptron and Precision of 0.829, Recall of 0.832, and F-measure of 0.817 using AdaBoost on a dataset containing 112 malignant and 298 benign lesion dermoscopic images.

A novel method for estimating a field of orientation distribution functions (ODF) from a given set of DW-MR images is presented. We model the ODF by Cartesian tensor basis using a parametrization that explicitly enforces the positive definite property to the computed ODF. The computed Cartesian tensors, dubbed Cartesian Tensor-ODF (CT-ODF), are symmetric positive definite tensors whose coefficients can be efficiently estimated by solving a linear system with non-negative constraints. Furthermore, we show how to use our method for converting higher-order diffusion tensors to CT-ODFs, which is an essential task since the maxima of higher-order tensors do not correspond to the underlying fiber orientations. We quantitatively evaluate our method using simulated DW-MR images as well as a real brain dataset from a post-mortem porcine brain. The results conclusively demonstrate the superiority of the proposed technique over several existing multifiber reconstruction methods.

We present a method to detect and classify the dermoscopic structure pigment network which may indicate early melanoma in skin lesions. We locate the network as darker areas constituting a mesh, as well as lighter areas representing the 'holes' which the mesh surrounds. After identifying the lines and holes, 69 features inspired by the clinical definition are derived and used to classify the network into one of two classes: Typical or Atypical. We validate our method over a large, inclusive, real-world dataset consisting of 436 images and achieve an accuracy of 82% discriminating between three classes (Absent, Typical or Atypical ) and an accuracy of 93% discriminating between two classes (Absent or Present).

We present a method for automatically segmenting skin lesions by initializing the random walker algorithm with seed points whose properties, such as colour and texture, have been learnt via a training set. We leverage the speed and robustness of the random walker algorithm and augment it into a fully automatic method by using supervised statistical pattern recognition techniques. We validate our results by comparing the resulting segmentations to the manual segmentations of an expert over 120 cases, including 100 cases which are categorized as difficult (i.e.: low contrast, heavily occluded, etc.). We achieve an F-measure of 0.95 when segmenting easy cases, and an F-measure of 0.85 when segmenting difficult cases.

Eye trackers have been used for medical image perception in radiology since the 1960s, with pioneering work by Harold Kundel and Dr. Calvin Nodine from the Department of Radiology, University of Pennsylvania Health System, Philadelphia [348, 457]. Initial studies involved radiologists viewing 2D X-ray images of the chest or breast on photographic film displayed on large “light boxes”. The aim of these early studies was to evaluate the use of image processing techniques, such as edge enhancement, to improve the visibility of lung tumours and hence reduce the error rates [351]. In order to make their experiments rigorous, Kundel and Nodine developed the technique of synthesising nodules photographically on otherwise normal chest images. The eye monitoring data provided understanding of the visual search process [346] and why errors occurred [349, 457]. These projects are motivated by the aim to provide eye-gaze driven decision support systems [123, 350, 456].

Images of the MRI signal intensity are normally constructed by taking the magnitude of the complex-valued data. This results in a biased estimate of the true signal intensity. We consider this as a problem of parameter estimation with a nuisance parameter. Using several standard techniques for this type of problem, we derive a variety of estimators for the MRI signal, some previously published and some novel. Using Monte Carlo experiments we compare the estimators we derive with others previously published. Our results suggest that one of the novel estimators we derive may strike a desirable trade-off between bias and mean squared error.

Most intensity-based fMRI registration methods do not account for the fact that the volumes being aligned may differ: one may have blood oxygen level dependent (BOLD) contrast while the other does not. Especially in least-squares registration, this can result in motion parameter errors that are correlated to the stimulus. An iterative technique to overcome this activation bias is proposed and analyzed. The method, using mostly off-the-shelf software, is able to find the least-squares solution to both the registration and activation detection problems simultaneously. The resulting motion parameters and activation maps are considerably more accurate, yielding two-thirds fewer false-positive and one-third fewer false-negative activations.

In this paper, we investigated the mechanism of dividing a 2D-object border into a set of local and global indentation and protrusion segments by extending the classic curvature scale-space filtering method. The resultant segments, arranged in hierarchical structures, can represent the object shape. Applying this technique, we derived a border irregularity measure for pigmented skin lesions. The measure correlated well with experienced dermatologists' evaluations and may be useful for measuring the malignancy of the lesion. Furthermore, we can use the method to discover all the bays in an aerial map.

The use of a cocktail of monoclonal antibodies in intraoperative radioimmunodetection of colorectal tumour has been evaluated in 14 patients. Eight patients had primary and six recurrent colorectal cancer. Three different monoclonal antibodies were used included B72.3, recognizing TAG72 antigen, FO23C3 and FO23C5 directed against two epitopes of CEA. The antibodies were labeled with 125I. During surgery the radioactive emission of tumour and normal tissue was evaluated by a gamma detecting probe. Ten patients showed positive intraoperative results and in particular 6 (75%) of the 8 patients with primary tumour and 4 (67%) of the 6 patients with recurrence, with a mean tumour to normal tissue ratio of 2.2. Intraoperative radioimmunodetection was instrumental in modifying the therapeutic approach in two patients with recurrent cancer but in none of the patients with a primary presentation. Immunohistochemical analysis, to evaluate TAG72 and CEA antigen expression, was performed in 13 cases. CEA antigen was expressed in 8/13 cases and TAG72 in 5/13. Thus the sensitivity was 38.5% and 61.5% with B72.3 and anti-CEA monoclonal antibodies respectively, while in combination the sensitivity was 71.4%. The use of a cocktail of different antibodies improved the sensitivity of intraoperative radioimmunodetection of colorectal tumours, when compared with a single antibody injection. This approach might improve the clinical use of radioimmunodetection.

This paper describes a robust fully automatic method for segmenting the brain from head MR images, which works even in the presence of RF inhomogeneities. It has been successful in segmenting the brain in every slice from head images acquired from three different MRI scanners, using different resolution images and different echo sequences. The three-stage integrated method employs image processing techniques based on anisotropic filters, „snakes“ contouring techniques, and a-priori knowledge. First the background noise is removed leaving a head mask, then a rough outline of the brain is found, and finally the rough brain outline is refined to a final mask.