Jonathan Rees

A Color and Texture Based Hierarchical K-NN Approach to the Classification of Non-melanoma Skin Lesions

This chapter proposes a novel hierarchical classification system based on the K-Nearest Neighbors (K-NN) model and its application to non-melanoma skin lesion classification. Color and texture features are extracted from skin lesion images. The hierarchical structure decomposes the classification task into a set of simpler problems, one at each node of the classification. Feature selection is embedded in the hierarchical framework that chooses the most relevant feature subsets at each node of the hierarchy. The accuracy of the proposed hierarchical scheme is higher than 93 % in discriminating cancer and potential at risk lesions from benign lesions, and it reaches an overall classification accuracy of 74 % over five common classes of skin lesions, including two non-melanoma cancer types. This is the most extensive known result on non-melanoma skin cancer classification using color and texture information from images acquired by a standard camera (non-dermoscopy).

A query-by-example content-based image retrieval system of non-melanoma skin lesions

This paper proposes a content-based image retrieval system for skin lesion images as a diagnostic aid. The aim is to support decision making by retrieving and displaying relevant past cases visually similar to the one under examination. Skin lesions of five common classes, including two non-melanoma cancer types are used. Colour and texture features are extracted from lesions. Feature selection is achieved by optimising a similarity matching function. Experiments on our database of 208 images are performed and results evaluated.

The Importance of a Full Clinical Examination: Assessment of Index Lesions Referred to a Skin Cancer Clinic Without a Total Body Skin Examination Would Miss One in Three Melanomas

Traditional clinical teaching emphasises the importance of a full clinical examination. In the clinical assessment of lesions that may be skin cancer, full examination allows detection of incidental lesions, as well as helping in the characterisation of the index lesion. Despite this, a total body skin examination is not always performed. Based on two prospective studies of over 1,800 sequential patients in two UK centres we show that over one third of melanomas detected in secondary care are found as incidental lesions, in patients referred for assessment of other potential skin cancers. The majority of these melanomas occurred in patients whose index lesion turned out to be benign. Alternative models of care--for instance some models of teledermatology in which a total body skin examination is not performed by a competent practitioner--cannot be considered equivalent to a traditional consultation and, if adopted uncritically, without system change, will likely lead to melanomas being missed.

Novice identification of melanoma: not quite as straightforward as the ABCDs.

The ABCD mnemonic to assist non-experts diagnosis of melanoma is widely promoted; however, there are good reasons to be sceptical about public education strategies based on analytical, rule-based approaches--such as ABCD (i.e. Asymmetry, Border Irregularity, Colour Uniformity and Diameter). Evidence suggests that accurate diagnosis of skin lesions is achieved predominately through non-analytical pattern recognition (via training examples) and not by rule-based algorithms. If the ABCD are to function as a useful public education tool they must be used reliably by untrained novices, with low inter-observer and intra-diagnosis variation, but with maximal inter-diagnosis differences. The three subjective properties (the ABCs of the ABCD) were investigated experimentally: 33 laypersons scored 40 randomly selected lesions (10 lesions × 4 diagnoses: benign naevi, dysplastic naevi, melanomas, seborrhoeic keratoses) for the three properties on visual analogue scales. The results (n = 3,960) suggest that novices cannot use the ABCs reliably to discern benign from malignant lesions.

Dermatology undergraduate skin cancer training: a disconnect between recommendations, clinical exposure and competence

BackgroundSkin cancers are the most common malignancies in Caucasian populations. Non-specialists are responsible for the initial assessment of skin lesions and are required to act as the gatekeepers to dermatological cancer services in many healthcare systems. The majority of such physicians receive very limited formal undergraduate or postgraduate dermatology training. The British Association of Dermatologists (BAD) has produced guidelines that list the lesions that students should be able to diagnose on graduation and the majority of UK medical schools’ operate curricula in keeping with these. There is, however, virtually no evidence as to whether these competencies are being achieved. We set out to determine students’ competence at skin lesion diagnosis and to quantify their clinical exposure to examples of such lesions during their dermatology attachment.MethodsThree linked studies were undertaken. In the first, students’ competence was tested by randomized slideshows of images containing the 16 lesions recommended in the UK guidelines. Students’ accuracy was tested at the beginning (Day 1) and end (Day 10) of their clinical placement, with a random sample of students retested 12u2009months later. Secondly, students’ exposure to these lesions was recorded during their attachments. Finally a survey of the additional dermatological resources used by the students was undertaken.ResultsStudy 1: Students’ diagnostic accuracy increased from 11% on Day 1 to 33% on Day 10 (effect size +2.72). After 12u2009months half of this effect had disappeared and the students accuracy had dropped to 24%. Study 2: Students’ exposure to the recommended lesions was poor with 82% not even witnessing a single example of each of the 3 major skin cancers. Despite these measurements, only a minority of students reported that they were not confident at diagnosing skin tumours. Study 3: The majority of students use additional resources to supplement their learning.ConclusionsIn the light of what we know about learning in dermatology, our data would suggest, that the current (traditional) undergraduate attachment is inadequate to meet the UK recommendations for graduate competence. As well as critically examining the basis for these recommendations, we need more empirical data on student performance and exposure, in order to improve teaching and learning.

Estimating the ground truth from multiple individual segmentations incorporating prior pattern analysis with application to skin lesion segmentation

Having ground truth is critical for evaluating segmentation algorithms and estimating the ground truth from a collection of manual segmentations remains a hard problem. A proper estimation approach should take into account and compensate for the inter-rater variation. In this paper, we conduct an analysis of manual segmentations in order to have a better understanding of the pattern of the variation and investigate whether incorporating such pattern information will improve the ground truth estimation. We propose a level-set based approach that solves the ground truth estimation in a probabilistic formulation. The prior pattern information is incorporated into the estimation model by adding a specially designed term in the energy function. Experiments on both synthetic and real data show that this prior information helps to find a more accurate estimate of the ground truth.

Utility of Non-rule-based Visual Matching as a Strategy to Allow Novices to Achieve Skin Lesion Diagnosis

Non-analytical reasoning is thought to play a key role in dermatology diagnosis. Considering its potential importance, surprisingly little work has been done to research whether similar identification processes can be supported in non-experts. We describe here a prototype diagnostic support software, which we have used to examine the ability of medical students (at the beginning and end of a dermatology attachment) and lay volunteers, to diagnose 12 images of common skin lesions. Overall, the non-experts using the software had a diagnostic accuracy of 98% (923/936) compared with 33% for the control group (215/648) (Wilcoxon pu2009<u20090.0001). We have demonstrated, within the constraints of a simplified clinical model, that novices diagnostic scores are significantly increased by the use of a structured image database coupled with matching of index and referent images. The novices achieve this high degree of accuracy without any use of explicit definitions of likeness or rule-based strategies.

Depth Data Improves Skin Lesion Segmentation

This paper shows that adding 3D depth information to RGB colour images improves segmentation of pigmented and non-pigmented skin lesion. A region-based active contour segmentation approach using a statistical model based on the level-set framework is presented. We consider what kinds of properties (e.g., colour, depth, texture) are most discriminative. The experiments show that our proposed method integrating chromatic and geometric information produces segmentation results for pigmented lesions close to dermatologists and more consistent and accurate results for non-pigmented lesions.

Content-based image retrieval of skin lesions by evolutionary feature synthesis

This paper gives an example of evolved features that improve image retrieval performance. A content-based image retrieval system for skin lesion images is presented. The aim is to support decision making by retrieving and displaying relevant past cases visually similar to the one under examination. Skin lesions of five common classes, including two non-melanoma cancer types, are used. Colour and texture features are extracted from lesions. Evolutionary algorithms are used to create composite features that optimise a similarity matching function. Experiments on our database of 533 images are performed and results are compared to those obtained using simple features. The use of the evolved composite features improves the precision by about 7%.

Non-melanoma skin lesion classification using colour image data in a hierarchical K-NN classifier

This paper presents an algorithm for classification of non-melanoma skin lesions based on a novel hierarchical K-Nearest Neighbors (K-NN) classifier. The K-NN classifier is simple, quick and effective. The hierarchical structure decomposes the classification task into a set of simpler problems, one at each node of the classification. Feature selection is embedded in the hierarchical framework that chooses the most relevant feature subsets at each node of the hierarchy. Colour and texture features are extracted from skin lesions. The accuracy of the proposed hierarchical scheme is higher than 93% in discriminating cancer and pre-malignant lesions from benign lesions, and it reaches an overall classification accuracy of 74% over five common classes of skin lesions, including two non-melanoma cancer types. This is the most extensive published result on non-melanoma skin cancer classification from colour images acquired by a standard camera (non-dermoscopy).