II.2 Automatic Diagnosis

(1)

II.2.1 Basic Aims and Requirements Diagnosis of melanocytic skin tumours is tradi- tionally based on the subject of evaluation of images by an expert in the discipline. Despite increasing refinement of subjective criteria – in- troducing some degree of “objectivity” – there is a continuous demand for truly objective diagnostic features. This means features independent of the subjective judgement of a human observer, or, more drastically, features created and interpreted by a machine.

The essential prerequisite for any approach for an automatic melanocytic lesion diagnosis is a set of digital data which can be used for automated analysis. in most approaches, this data set represents a digitized image which may have been acquired by any number of methods, rang- ing from clinical photography to three-dimensional reconstruction of confocal laser scanning

Automatic Diagnosis

Josef Smolle

II.2

Contents

II.2.1 Basic aims and requirements . . . .47

II.2.2 Clinical images . . . .47

II.2.3 Digital Dermoscopy . . . .48

II.2.4 other Methods . . . .48

II.2.4.1 Spectral analysis . . . .48

II.2.4.2 in-Vivo Confocal laser Scanning Microscopy . . . .48

II.2.4.3 optical Coherence Tomography . . . .49

II.2.4.4 levels of Evidence . . . .49

II.2.4.5 Conclusion . . . .50

references . . . .50

microscopic images. The entire undertaking may serve two different goals: on the one hand, digital data processing may be used to enhance visually recognizable criteria – which finally are still evaluated by a human observer. on the other hand, digital data processing should directly result in a diagnostic suggestion generated by the machine independently of the human observer. The latter approach is the more fascinat- ing one, although it has not yet revealed its full potential.

II.2.2 Clinical Images

Some efforts have been made to use clinical images for automatic diagnosis. Usually colour is taken as the main source of data, although colour itself is hardly a reliable parameter. Usually, some kind of distribution (or texture) analysis has to be considered. Chen et al., for example, showed that not the presence of melanoma-spe- cific colour pixels per se, but colour clustering is the more reliable feature [3]. a preliminary report by Manousaki and coworkers [15] deals with subtle colour features: intensity values of each of the three colour channels were plotted as the third dimension of the plane clinical pho- tograph, and the surface of the plot was anal- ysed by methods of fractal mathematics includ- ing fractal dimensionality and lacunarity. a major problem is often the identification of rel- evant features. Chang and colleagues therefore proposed a systematic heuristic approach to feature selection particularly applicable to clinical images [2].

Clinical images of individual lesions are usually inferior to dermoscopic images with respect to automated classification [22]. Clinical imag-

(2)

II.2

es, however, are gaining some importance as whole-body screening tools. The detection and demarcation of pigmented lesions alone is a sophisticated task [9] which is an essential prerequisite for any subsequent detailed analysis and automatic diagnosis.

II.2.3 Digital Dermoscopy

at present, dermoscopic digital images seem to be the most promising source for automatic melanoma diagnosis. This may be due in part to the fact that standardization is more easily achieved than in clinical imaging, and due in part to the higher magnification and pixel resolution. The potential target structures are mani- fold: Stoecker and coworkers identified asymmetric structureless areas (“blotches”) [28], whereas Seidenari et al. based their classification algorithm on average colour values within square pixel blocks [25]. Following the method of a preliminary study by Kahofer et al. [13], Gerger and coworkers used tissue counter analysis [27] to diagnose melanoma in dermoscopy images [6].

rubegni et al. [23] and oka et al. [18] developed advanced diagnostic systems based on dermoscopic images, one as a built-in-module of a digital dermoscopy device, and the other as a classification program accessible via the internet.

Particular emphasis has been put on the pos- sibility of sequential dermoscopic images. Vi- sual analysis of consecutive images taken a few months apart has been shown to increase the proportion of true melanomas within the set of lesions which were finally excised [10], and fa- cilitated the detection of melanomas which did not display the usual diagnostic features (so- called featureless melanomas) [14]. automatic comparison of sequential images would be a valuable undertaking in the future.

II.2.4 Other Methods

There are a growing number of imaging methods for melanocytic lesions. although most of them have not yet been proven to be reliable for automatic diagnostic procedures, there are promising preliminary results.

II.2.4.1 Spectral Analysis

Spectral analysis is a method beyond simple three-channel colour analysis. Spectral analysis creates a three-dimensional data cube of two- dimensional images, with each plane of the cube representing a particular wavelength [4]. De- pending on the attempted wavelength resolution, up to several hundred two-dimensional images would be possible. Usually, however, analysis is limited to a small number of wave- lengths which had turned out to be of discrimi- natory power. in 2001, Farkas and Becker re- ported automatic detection of the melanoma component in a complex melanocytic skin lesion based on spectral analysis. Pseudo-colour images clearly denoted the malignant portion of the lesion [4]. Spectral intracutaneous analysis (Sia) is based on eight narrow-width wavelength images between 400 and 1000 nm and facilitates the demonstration of melanin, hae- moglobin and collagen within skin lesions [16].

Melanomas usually present peculiar patterns which might be suitable for automatic diagnosis. Murphy et al. [17] applied fibre-optic diffuse reflectance spectroscopy to melanocytic skin lesions and found a remarkably high degree of diagnostic accuracy. another highly sophisticated approach is raman spectroscopy. This type of laser-induced spectral analysis is based on molecular vibrations and therefore represents to some degree the chemical composition of a lesion. Gniadecka et al. [7] applied this method to freshly excised tissue specimens ob- tained by punch biopsy and achieved a diagnostic accuracy comparable to automatic dermoscopy analysis.

II.2.4.2 In-Vivo Confocal

Laser Scanning Microscopy in-vivo confocal laser scanning microscopy facilitates non-invasive examination of superficial skin layers at the cellular level [5]. Qualitative and semiquantitative diagnosis of melanoma is largely based on the architectural arrangement of keratinocytes in the spinous layer and on the size, shape and distribution of pigmented cells [5]. Digital image processing has been used to

(3)

enhance visibility of criteria [8]. recently, ono et al. introduced three-dimensional reconstruction of in-vivo confocal laser scanning microscopy image stacks [19]. Experimental studies examining the diagnostic significance of tissue counter and wavelet analysis [30] are on the way (M. Wiltgen et al., pers. commun.).

II.2.4.3 Optical Coherence Tomography

optical coherence tomography provides verti- cal sections through the skin [1]. resolution is somewhere between ultrasound imaging, on the one hand, and confocal laser scanning microscopy, on the other. it has been used to evaluate histological features underlying certain dermoscopic criteria. Since the method, in addition to imaging, provides quantitative information on the refractive index and the scattering coeffi- cient of the lesion, it might contribute to automatic diagnosis.

initially, ultrasound imaging was used to de- termine lesion thickness in melanomas in which the diagnosis had been determined by other criteria [24]. More sophisticated technical variants of the method, however, can produce quantitative data related to tissue architecture. rallan et al. assessed attenuation characteristics of melanocytic skin lesions using reflex transmission imaging [20]. in a pilot study, high diagnostic accuracy was achieved when these data were ac- companied by digital analysis of clinical images.

Subsequently, the authors demonstrated the diagnostic utility of the quantitative parameters per se [21].

another approach to lesion interpretation is surface scanning. The method is based on the observation that the skin surface of malignant lesions may differ significantly from those of normal skin. Mathematical morphology facilitates the extraction of skin lines from common white-light clinical images. The skin-line texture seems to hold promising diagnostic information [26].

There is an increasing body of methods which reveal functional features – instead of, or in addition to, image data. Stücker and coworkers showed an increase of blood flow by laser Dop-

pler flowmetry in malignant, as compared with benign, melanocytic lesions [29]. This increase in blood flow was more pronounced than the increase in vessel density at the histological level – clearly indicating a functional difference in addition to the morphological finding of neo- vascularization. This finding may in some way correspond to increased po2 levels in melanomas compared with nevi. po2 imaging can be performed with the SkinCam system [12]. an astounding approach is the measurement of electric tissue impedance. impedance differs significantly between benign and malignant lesions, and combining impedance measurements with digital image analysis provides high diagnostic accuracy [11].

II.2.4.4 Levels of Evidence

in view of the impressive body of studies – just a few of them have been cited in this chapter – one may wonder why automatic melanoma diagnosis has not yet become a routine application. The studies are technically sound and usually pro- vide a sensitivity close to 95 or 100%, together with a specificity which is considerably higher than that of more conventional procedures. one has to consider, however, that the level of evidence may still be insufficient to justify routine application in everyday practice. The possible levels of evidence may be classified as follows:

1. Statistical difference between two diagnostic groups by univariate or multivariate analysis

2. Classification generated in, and applied to, a single selected set of lesions

3. Classification generated in a learning set and applied to an independently selected test set

4. application to an independent, non- selected (e.g. randomly or consecutively sampled) test set

5. application to an independent, non- selected, consecutively sampled test set acquired in different institutions 6. application in everyday routine work in

numerous institutions

(4)

II.2

Many studies start with level i and reach level ii or even level iii. in contrast, the number of studies providing levels iV, V or Vi is small.

Particularly the application to consecutively sampled lesions in routine diagnostic work is still quite uncommon, but is an essential prerequisite for general acceptance as a truly useful diagnostic tool.

II.2.4.5 Conclusion

automatic analysis of digital dermoscopic images is at the threshold of routine application as an additional diagnostic procedure for melanocytic skin lesions. Confocal laser scanning microscopy and spectral imaging techniques hold promise for the future. an automatic procedure will add information to the diagnostic decision process similarly as an expert second opinion.

Ultimately, the number of unnecessarily re- moved benign lesions will be reduced.

C Core Messages

■ automatic diagnostic procedures of melanocytic skin lesions can be applied to a broad range of morphological and functional imaging techniques.

■ at present, automatic analysis of digital dermoscopy images is the most advanced approach.

■ in-vivo confocal laser scanning microscopy and spectral imaging techniques hold considerable promise for the future.

■ an automatic diagnostic procedure, once, will serve as an expert second opinion.

References

1. (2005) Possible histopathologic correlates of dermoscopic features in pigmented melanocytic lesions identified by means of optical coherence tomography. Exp Dermatol 14: 56–59

2. Chang Y, Stanley rJ, Moss rH, Van Stoecker W (2005) a systematic heuristic approach for feature selection for melanoma discrimination using clinical images. Skin res Technol 11: 165–178

3. Chen J, Stanley rJ, Moss rH, Van Stoecker W (2003) Colour analysis of skin lesion regions from melanoma discrimination in clinical images. Skin res Technol 9: 94–104

4. Farkas Dl, Becker D (2001) applications of spectral imaging: detection and analysis of human melanoma and its precursors. Pigment Cell res 14: 2–8 5. Gerger a, Koller S, Weger W, richtig E, Kerl H, Sa-

monigg H, Krippl P, Smolle J (2006) Sensitivity and specificity of confocal laser-scanning microscopy for in vivo diagnosis of malignant skin tumors.

Cancer 107: 193–200

6. Gerger a, Stolz W, Pompl r, Smolle J (2003) automated epiluminescence microscopy-tissue counter analysis using CarT and 1-NN in the diagnosis of melanoma. Skin res Technol 9: 105–110

7. Gniadecka M, Philipsen Pa, Sigurdsson S, Wessel S, Nielsen oF, Christensen DH, Hercogova J, rossen K, Thomsen HK, Gniadecki r, Hansen lK, Wulf HC (2004) Melanoma diagnosis by raman spectroscopy and neural networks: structure alterations in proteins and lipids in intact cancer tissue. J invest Dermatol 122: 443–449

8. Gruber MJ, Wackernagel a, richtig E, Koller S, Kerl H, Smolle J (2005) Digital image enhancement for in vivo laser scanning microscopy. Skin res Tech- nol 11: 248–253

9. Guillard G, lagarde JM (2005) Skin lesions segmen- tation and quantification from 3D body’s models.

Skin res Technol 11: 123–131

10. Haenssle Ha, Krueger U, Vente C, Thoms K-M, Bertsch HP, Zutt M, rosenberger a, Neumann C, Emmert S (2006) results from an observational trial: digital epiluminescence microscopy follow- up of atypical Nevi increases the sensitivity and the chance of success of conventional dermoscopy in de- tecting melanoma. J invest Dermatol 126: 980–985 11. Har-Shai Y, Glickman Y, Siller G, Mcleod r, Topaz

M, Howe C, Ginzburg a, Zamir B, Filo oP, Kenan GP, Ullmann Y (2005) Electrical impedance scanning for melanoma diagnosis: a validation study.

Plast reconstr Surg 116: 782–790

12. Hartmann P, Ziegler W, lübbers DW (1996) Fluo- rescence life time imaging of the skin po2: instru- mentation and results. adv Exp Med Biol 428: 605–

611

(5)

13. Kahofer P, Hofmann-Wellenhof r, Smolle J (2002) Tissue counter analysis of dermatoscopic images of melanocytic skin tumors: preliminary findings.

Melanoma res 12: 71–75

14. Kittler H, Guitera P, riedl E, avramidis M, Teban l, Fiebiger M, Weger ra, Dawid M, Menzies S (2006) identification of clinically featureless incipient melanoma using sequential dermoscopy imaging. arch Dermatol 142: 1113–1119

15. Manousaki aG, Manios aG, Tsompanaki Ei, Pan- ayiotides JG, Tsiftsis DD, Kostaki aK, Tosca aD (2006) a simple digital image processing system to aid in melanoma diagnosis in an everyday melanocytic skin lesion unit. a preliminary report . int J Dermatol 45: 402–410

16. Moncrieff M, Cotton S, Claridge E, Hall P (2002) Spetrophotometric intracutaneous analysis: a new technique for imaging pigmented skin lesions. Br J Dermatol 146: 448–457

17. Muryph BW, Webster rJ, Turlach Ba, Quirk CJ, Clay CD, Heenan PJ, Sampson DD (2005) Towards the discrimination of early melanoma from common and dysplastic nevus using fiber optic diffuse reflectance spectroscopy. J Biomed optics 10:

064020

18. oka H, iyatomi H, Hashimoto M, Tanaka M (2006) reply to “Digital dermoscopy analysis and internet- based program for discrimination of pigmented skin lesion dermoscopic images”. Br J Dermatol 154: 570–571

19. ono i, Sakemoto a, ogino J, Kamyia T, Yamashita T, Jimbow K (2006) The real-time, three-dimensional analyses of benign and malignant skin tumors by confocal laser scanning microscopy. J Dermatol Sci 43: 135–141

20. rallan D, Dickson M, Busch Nl, Harland CC, Mortimer P, Bamber JC (2006) High-resolution ultrasound reflex transmission imaging and digital photgraphy: potential tools for the quantitative assessment of pigmented lesions. Skin res Technol 12:

50–59

21. rallan D, Bush Nl, Bamber JC, Harland CC (2006) Quantitative discrimination of pigmented lesions using three-dimensional high-resolution ultrasound reflex transmission imaging. J invest Derma- tol 127: 189–195

22. rosado B, Menzies S, Harbauer a, Pehamberger H, Wolff K, Binder M, Kittler H (2003) accuracy of computer diagnosis of melanoma. a quantitative meta-analysis. arch Dermatol 139: 361–367 23. rubegni P, Burroni M, Sbano P, andreassi l (2005)

Digital dermoscopy analysis and internet-based program for discrimination of pigmented skin lesion dermoscopic images. Br J Dermatol 152: 395–396 24. Salmhofer W, rieger E, Soyer HP, Smolle J, Kerl H

(1996) influence of skin tension and formalin fixa- tion on sonographic measurement of tumor thickness. J am acad Dermatol 34: 34–39

25. Seidenari S, Pellacani G, Grana C (2005) Pigment distribution in melanocytic lesion images: a digital parameter to be employed for computer-aided diagnosis. Skin res Technol 11: 236–241

26. She Z, Fish PJ (2003) analysis of skin line pattern for lesion classification. Skin res Technol 9: 73–80 27. Smolle J (2000) Diagnostic assessment of cutaneous

melanoma and common nevi using tissue counter analysis. analyt Quant Cytol Histol 22: 299–306 28. Stoecker VW, Gupta K, Stanley rJ, Moss rH, Shres-

tha B (2005) Detection of asymmetric blotches (asymmetric structureless areas) in dermoscopy images of malignant melanoma using relative color.

Skin res Technol 11: 179–184

29. Stücker M, Springer C, Paech V, Hermes N, Hoff- mann M, altmeyer P (2006) increased laser Dop- pler flow in skin tumors corresponds to elevated vessel density and reactive hyperemia. Skin res Technol 12: 1–6

30. Wiltgen M, Gerger a, Wagner C, Bergthaler P, Smolle J (2004) Evaluation of the influence of image compression on the automatic discrimination of histological images of skin lesions. Meth inform Med 43: 141–149