18 March 1998
British Institute of Radiology, 36 Portland Place, London
Chair: Ela Claridge, School of Computer Science, The University of Birmingham
PROGRAMME
Moles and machine vision
Per Hall (Department of Plastic Surgery, Addenbrooke's Hospital, Cambridge)
Per N Hall FRCS(Plast)* and Ela Claridge**
* Addenbrooke's Hospital, Hills Road, Cambridge CB2 2QQ
** School of Computer Science, The University of Birmingham, Birmingham B15 2TT
E.Claridge@cs.bham.ac.uk
It should be possible for all doctors to recognise what might be a cutaneous malignant melanoma, even if they don't know for sure, since early diagnosis can lead to cure whilst disseminated disease is generally fatal. Worldwide the incidence of this disease continues to rise at approximately 10% per annum. In the UK the incidence in 1994 was 7.8 per 100,000 for invasive melanoma in men and 12.3 per 100,000 in women. In is worth comparing this with an incidence of 3.5 per 100,000 in men and 6.8 per 100,000 in women in 1979. The average family doctor will see one melanoma every 4 years of their practice life. We cannot expect GPs to know the diagnosis of every pigmented lesion they see since even experienced clinicians can get the precise diagnosis wrong. Can machine vision help sort our worrying ones from the friendly ones? Can it be used to guide inexperienced doctors towards the features which can help make diagnosis and teach doctors in training? Can analysis of the lesion by digital image processing alone ever be enough to reliably never miss a malignant melanoma?
Clinical examples of benign and malignant skin lesions will be
shown. A short summary of the work resulting from a collaboration between a clinician and
a department of computer science will be given.
John Curnow*, Peter Kersey** and Stephen Brown*.
* Department of Medical Physics, Derriford Hospital Plymouth Devon
** Department of Dermatology, Derriford Hospital Plymouth Devon
john.curnow@phnt.swest.nhs.uk
The problem of providing an easily available quality diagnosis of pigmented skin lesions has been investigated by a number of groups over the last ten or more years. Several different approaches have been tried and claims for each have been made. We have reviewed the literature up to early 1997 and identified the important approaches made and applied a consistent approach to the analysis of the results to allow a comparison where possible of the effectiveness of each. The methods can be split into three groups with several different methods in each group. Within the field of computer image processing several areas can be identified as having been investigated without full success.
Figure 1 below (to be inserted) shows the comparison of the published results for each method, this shows that Clinical diagnosis by an expert is as good as Dermatoscopy by an expert and that other methods are still not as good as either of these. In general however all new methods have only been tried on limited data sets, often with some pre-sorting to help the identification and have only been tested by the experts who have developed them. With Dermatoscopy where non experts have been tried it has been shown that the method reduces the effectiveness of the non expert rather then enhancing it.
To date no system has been developed and proved clinically that can
provide a tool for primary care physicians to promote their expertise in categorising a
pigmented skin lesion to that of an expert.
W.E. Denton, A.W.G. Duller and P.J. Fish
School of Electronic Engineering and Computer Systems, University of Wales, Bangor,
Dean St., Bangor, Gwynedd, LL57 1UT
andy@sees.bangor.ac.uk
The accurate location of the boundary of skin lesions is an
important first step in the automatic diagnosis of malignant melanoma. The use of standard
edge detectors for skin lesion boundary detection from grey scale images has serious
shortcomings since those giving sufficiently detailed borders are sensitive to spurious
small scale structure elsewhere in the image while those which ignore small scale
structure involve low-pass filtering which loses border detail. Edge focusing uses edge
detectors of progressively smaller scale, each focusing only in the region of the boundary
estimated by the previous detector. This results in a series of boundary position
estimates of increasingly finer scale while eliminating spurious edges resulting from
noise and other image detail. The selection of the "best fit" boundary is made
by comparison of the contrast between narrow regions just inside and outside the boundary
or from the variance of image intensity in the latter region.
A.J. Round, A.W.G. Duller, P.J.Fish.
School of Electronic Engineering and Computer Systems, University of Wales, Bangor,
Dean St., Bangor, Gwynedd, LL57 1UT
andy@sees.bangor.ac.uk
This paper describes a method of distinguishing between early malignant melanoma and benign moles by examining texture on an image of the lesion. Normal skin (except the palms and soles) shows a pattern of lines cris-crossing its surface and this pattern tends to be disrupted by cancerous melanoma but remains over many benign forms.
Skin patterning is a macroscopic texture composed of fine linear elements. This texture is poorly described by standard definitions of texture and poorly detected by existing techniques. In the method described here skin line patterning is detected through a process which looks at small patches spaced equally across the image and evaluates their linear self-similarity - using the width of the autocorrelation function - as a function of patch orientation. Regions which exhibit skin patterning result in similar width- versus-angle profiles for neighbouring patches whereas no such similarity is found in areas where the patterning is disrupted. Interpretation of the profile images for the classification of the lesions is then addressed.
Three methods are used to classify the output image of profiles from the skin patterning detector, one is based on a self organising neural network (SOFM), the second uses the local variance and the last is a region based agglomerative clustering technique. A measure based on the relationship between the classification results and an intensity based segmentation is calculated and this represents the disruption of the skin line patterning (i.e. if the texture based classification reveals the lesion then the skin structure has been damaged by the lesion). A set of images containing a variety of malignant and non-malignant lesions are analysed. The computed textural disruption figure is compared to both the histological diagnosis and to a visual estimate of patterning disruption for each image. Comparative results are given for the discriminative ability obtained via each of the classification methods.
The degree of skin pattern disruption is being considered as only one of a number of lesion descriptors forming a vector of feature estimates to be used as an aid to diagnosis. Nevertheless, initial results suggest that a high degree of lesion-type separation is possible using only skin line analysis with all the methods investigated.
Problems arise as a result of loss of focus due to limb curvature
leading to apparent loss of skin line at image boundaries, hairs and crusting on benign
lesions. These cause the method to fail and overcoming these difficulties is the subject
of further research.
A Babaramo, J Cook, C Dewdney, G Eley, DR Prytherch
Division of Physics, University of Portsmouth,
St Michael's Building, White Swann Road, Portsmouth PO1 2DT
chris.dewdney@port.ac.uk
The aim of the research described here is to improve the accuracy
and efficiency of the pigmented skin-lesion diagnostic process in both primary and
secondary health care environments. The basis of the research is the use of digital image-
processing techniques to identify and quantify those machine-measurable features of
skin-lesion images that are relevant for diagnostic and screening purposes. Our results
show that whereas quantifying the purely geometric features of lesions is of limited use,
quantifying aspects of their coloration yields highly sensitive and specific measures. The
data set for this study consisted of optical images of 42 malignant melanoma and 75 benign
lesions all histologically confirmed after excision and a further 39 benign (clinically
diagnosed but non- excised) lesions. In this data set the average red and average green
components of the regions of clinical interest were found to yield measures 81% sensitive
and 87% specific for the red and 81% sensitive and 85% specific for the green,
(significantly better than some reported clinical diagnostic rates). In the non-excised
set the agreement between this measure and the clinical diagnosis was 100%.
Jon Morris Smith Formerly at School of Computer
Science, The University of Birmingham,
Birmingham B15 2TT
jon.morris-smith@gs.com
Frequently image-based diagnostic tools are based on the premise
that results returned by algorithms accurately reflect the assessment of features by
clinicians. If a technique is to act as an aid to clinical diagnosis it is important that
it accurately reflects feature assessment by expert clinicians. One such technique is the
application of fractal-based measures to assess lesion border irregularity. This work
shows how fractal-dimension corresponds closely to the expert assessment of lesion-border
irregularity in isolation. The work goes on to show that the assessment of border
irregularity is affected by the presence of textures within the outlines. This result
indicates that, when applying a metric to a lesion feature, care must be taken to ensure
that any feature interactions are taken into account if the measure is to accurately
reflect the assessment by clinicians: it is not necessarily a valid assumption that a
metric which is good for measuring a feature in isolation can equally be applied to
"real" data.
Symon Cotton
School of Computer Science, The University of Birmingham,
Birmingham B15 2TT
S.D.Cotton@cs.bham.ac.uk
A non-invasive method for assisting in the diagnosis of malignant melanoma by quantifying diagnostically relevant histological information from the analysis of one colour and two infrared images of a lesion is presented. In particular, it is shown that the presence of melanin within the dermis can be detected when its invasion from the dermo-epidermal junction is 0.02mm.
Through development of an optical image formation model of human
skin it is shown that all normal skin colours lie on a two-dimensional surface within a
three-dimensional colour-space. In contrast, colour co-ordinates corresponding to the
penetration of melanin into the dermis deviate from this surface along characteristic
paths and thus can be identified. This solution, however, can not be directly applied
because the colours resulting from a decrease in the thickness of the top layer of the
dermis, the papillary dermis, can occupy the same position in the colour space as colours
resulting from melanin descent. This ambiguity can be resolved by transforming the
measured colour co-ordinates to compensate for the thickness of the papillary dermis. It
is shown that this thickness can be obtained through the analysis of a pair of infrared
images. It is further shown that the amount of epidermal melanin, dermal blood and
thickness of the papillary dermis can also be quantified.
JC Bamber
Institute of Cancer Research and Royal Marsden Hospital NHS Trust,
Downs Road, Sutton, Surrey, SM2 5PT
jeff@icr.ac.uk
The rising incidence of malignant melanoma, combined with evidence that dermatologists and general practitioners have varying success in diagnosing the disease, suggest the need for an objective aid to diagnosis. We have investigated both optical reflectance spectrophotometry and quantitative high-resolution ultrasound methods, with a view to devising easy to use methods that will eventually reduce the number of unnecessarily excised benign lesions.
The optical reflectance characteristics of 260 pigmented skin lesions have been documented for the wavelength range 320 to 1100 nm. Multiple discriminant analysis and an artificial neural network were used to evaluate the diagnostic performance of an objective system based on the spectral features determined to have the best power to discriminate between benign and malignant lesions. The artificial neural network achieved the best diagnostic performance, discriminating with sensitivity 100% and specificity 65% between cases of malignant melanoma and benign naevi that had previously been marked for excision because of clinical suspicion of malignancy. Semi-quantitative measurements of histological features likely to be optically important were found to be related in a complex manner to the spectral reflectance features but, via a Monte Carlo model, were able to predict the mean optical reflectance of normal dermis, malignant melanoma and benign naevi.
A Cortex Dermascan 20 MHz ultrasound scanner interfaced to a computer was used to assess i) relative mean internal tumour echogenicity, ii) internal tumour echo heterogeneity, iii) degree of acoustic shadowing, iv) shadow heterogeneity, and v) entry echo enhancement. Results from 111 skin lesions demonstrated that, for 100% sensitivity, 93% specificity may be obtained distinguishing between melanoma and typical basal cell papilloma because of echo-shadowing and entry echo enhancement, but inclusion of atypical basal cell papilloma devoid of keratin reduced the specificity to 66%. Melanomas were not distinguishable from moles although there was evidence to suggest that this may be possible with further development.
To determine whether images of the size of the tissue's acoustical scattering elements might display tumours with a contrast greater than for B-scans, or provide quantitative diagnostic information, we developed methods for scatterer size imaging of skin tumours in vivo. Scatterer size, in models of the tissue's acoustical structure, was estimated by maximising correlation between theoretically predicted and measured frequency dependencies of the backscattering coefficient. Evaluation in vivo on 18 skin tumours, using RF data (11-23 MHz) from a Cortex DermascanTM, revealed that for melanoma and benign naevi both contrast and spatial resolutions of scatterer size images were inferior to those of B-scans. Average scatterer size, however, was smaller in naevi than melanoma, correlating with quantitative histopathological estimates following excision and providing new information that may contribute to differential diagnosis. Best performance occurred when the analysis was restricted to 17.5-23 MHz, indicating the need for higher frequencies studies.
Optical and ultrasound methods appear to complement each other and
are sufficiently specific that a combined approach is worthy of further development and
prospective evaluation in clinical practice. Such an approach should help to exclude the
possibility of malignancy in the majority of benign tumours that currently are most easily
confused with malignant melanoma.
John Curnow*, Peter Kersey** and Nick Outram***
* Department of Medical Physics, Derriford Hospital
** Department of Dermatology, Derriford Hospital
*** SECEE, Plymouth University, Plymouth, Devon
john.curnow@phnt.swest.nhs.uk
When reviewing previous published work on using image processing to support the diagnosis of Malignant Melanoma we recognised that each system had been tested by a relatively small and limited set of data. Often the data used by a development group will produce good results because it has been selectively collected for the project. For a database to successfully test a system it must have a full spread of types of lesions with representative numbers of absolute normals, absolute abnormals and uncertain cases. This type of database will also be the best for training any artificial intelligence systems that are developed.
We decided that before we get deeply involved in the development of a system for pigmented skin lesion diagnosis we would collect a database of good quality images and clinical data that provided representative data for all types of lesions in representative numbers.
To this end we have started by selecting a camera system proposed by Dr Hall from Cambridge to produce consistent reproducible photographic images. The images are all taken at a fixed camera skin distance and with consistent lighting and camera settings. The images include a grey scale held against the skin to allow control of image quality.
The database will consist of a photograph of each suspicious lesion sent to the Pigmented Skin Lesion Clinic at our hospital. Also where ever possible a very normal lesion on the same part of the patient is photographed. Clinical data is collected about skin type, occupational exposure to UV, recent previous suntan history, family history of Melanoma, when the lesion appeared, its size, elevation, any changes in shape, size or colour, if it itches, has a surrounding erythema, is ulcerated, if there has been any recently bleeding the approximate number of lesions the patient has and where on the body the photographed moles are. The clinical diagnosis is noted and when available the histology is added, there is also space for general comments.
The diagnosis is one difficult area. We could not arrange for histology to be carried out on each lesion. We have therefore decided to accept the Senior Clinicians diagnosis of normal lesions without histology and only have histology from lesions that are thought to be suspicious or are removed for other reasons. To ensure as far as possible that any misdiagnosed melanomas are followed up a special tracer label is placed in each patients notes when a photograph is taken and if later diagnosis shows a change, a new photograph and the new diagnosis will be added to the database and the previous pictures clinical data updated as initially diagnosed non-malignant later became malignant. It is hoped this will provide important information on the possible signs in the image and/or clinical history to point to later malignancy.
The proposal is to develop this database to 500 images in the first year and then to continue to collect to at least 2000 images. The data is to be held on a dedicated file server holding clinical research data and limited access made available via the internet using special password control. It is hoped this will become a standard database that can be used to test and compare all systems being developed for lesion diagnosis.