Noninvasive Skin Imaging

An earlier model of colour formation within normal human skin was extended to include architectural distortions associated with various pigmented skin lesions, including malignant melanoma. The extended ve-layer model makes it possible to derive parameters characterising the thickness and pigment composition of the skin layers from calibrated colour and infrared images of skin lesions. The extracted parameters can be used to reconstruct a full 3-dimensional model of the skin architecture which conveys information grossly comparable to that available through microscopical examination of biopsied skin tissue. This work forms a part of research at the University of Birmingham into developing theories and techniques which aim to aid clinicians in the early diagnosis of malignant melanoma. Our previous paper [2] presented a model of colour formation within normal human skin. The model is based on the Kubelka-Munk theory [5] of scattering and absorption within inhomogeneous materials and the physics pertaining to their colour properties. By considering the skin to be a layered construction of such materials, the stratum corneum, epidermis, papillary dermis and reticular dermis, and by exploiting the physics related to the optical interface between these layers, the model generates all possible colours occurring within normal human skin. In particular, the model predicts that all normal skin colours lie on a simple curved surface patch within a three-dimensional colour space bounded by two physiologically meaningful axes, one corresponding to the amount of melanin within the epidermis and the other to the amount of blood within the dermis. The model postulated that abnormal skin conditions would cause the skin colours to deviate from the predicted surface in the colour space. This paper exploits this idea further. In particular, it demonstrates that it is possible to derive detailed information about the internal skin architecture and composition grossly comparable to information available through the microscopical examination of tumour tissue but without incurring the problems inherent in obtaining a biopsy. The central premise explored here is, therefore, that as abnormal skin often has a di erent internal architecture to normal skin it is a fair proposition that the coloration may not be bounded to this surface; if this is true, then the nature of the deviation may yield important information about the skin architecture. To explore this the model presented in the previous paper was extended to predict the skin coloration associated with conditions where melanocytes penetrate into the dermis. These are commonly seen in such conditions as the benign blue nevus and invasive skin cancer, such as melanomas, often leading to the blue hues characteristic with these conditions [1]. The \form" and depth of this invasion is an extremely important diagnostic factor in determining the nature of a skin lesion and, if abnormal, the relevant treatment and prognosis; currently the only reliable method of obtaining this information is by biopsy. To account for the architectural distortion where dermal melanocytes occupy a region of the papillary dermis the model of normal skin needed to be extended to include additional layers. As can be seen from Figure 1, there are now ve distinct layers which can be combined to construct an extended model: a layer within the upper papillary dermis containing no melanin; a layer within the upper papillary dermis containing melanin; a layer within the lower papillary dermis containing melanin; and a layer within the lower papillary dermis containing no melanin and nally the epidermis. The extended model [3] speci es how the magnitude of any colour primary depends on these model parameters thus allowing exploration of the coloration expected for various skin conditions. A rst such \computational experiment" modelled in vitro dermal tissue, that is bloodless skin with the epidermis removed. This analysis showed that the coloration \does indeed move o the surface corresponding to normal skin" in situations where melanocytes have penetrated the dermis. Further analysis led to the conclusion that \In principle, therefore, if presented with a section of in vitro dermis it should be possible to assess both the presence, concentration and position of melanocytes by an examination of the coloration" [3]. For such an approach to be useful when applied to living, in vivo, tissue it was necessary to include the e ect of both melanin absorption within the epidermis and blood within the dermis. The result of this analysis showed that although the amount of blood could be quanti ed directly through an analysis of the remitted skin colour, when combined with dermal penetration of melanocytes, the amount of epidermal melanin could not. It was, however, possible to ascertain that \an amount" of dermal invasion had occurred, thus allowing areas of normal and abnormal skin to be segmented, but it was impossible to unravel the relative amounts of each. To gain insight into the internal architecture therefore requires that either the amount of epidermal melanin, penetration of dermal melanocytes or the concentration of these melanocytes are known. If one of these parameters can be speci ed then it becomes possible to disentangle the other two from a measure of the coloration. Through further research it is hoped to be able to measure one of these parameters directly for every point within a lesion image. At present, however, the following approach has been adopted. First, the normal areas of skin are identi ed that is areas with no dermal penetration of melanocytes and within these areas the amount of epidermal melanin is ascertained; this is then followed by an interpolation of these surrounding melanin values into the abnormal areas. This approach assumes that the amount of epidermal melanin does not change by a signi cant amount within these areas or if it does it varies 3 In normal skin melanocytes are restricted to the epidermis. in a predictable manner thus allowing the variation to be modelled. As described in [3] however it is necessary to ascertain one further parameter before such analysis can be undertaken. This parameter, the papillary dermal thickness, has a pronounced e ect on the light remitted from a skin lesion. Indeed the change in coloration due to a variation in this parameter is almost identical to that due to melanocytic descent thus leading to lesions only varying in this parameter being wrongly classi ed. As an example it is possible to nd malignant invasive melanomas with an identical coloration to that of simple warts. This result casts doubt on the e ectiveness of using purely colour information in the diagnosis of malignant melanoma and may o er itself as an explanation for the \moderate success achieved" by Umbaugh et al. [6] when they attempted to classify lesion types by an investigation of coloration. However, this is not to dismiss the usefulness of colour information when combined with other extracted lesion features such as that demonstrated by Dhawan and Sicsu [4] when they combined colour with a texture analysis; and Umbaugh et al. [7] when they applied arti cial intelligence techniques to variations in lesion colour. From the extended skin model the variation of remitted light with papillary dermal thickness can be ascertained. Therefore if this thickness were known for each image point it should be possible to calculate a transformation that adjusts the measured coloration to that of any speci ed papillary dermal thickness thus removing the metameric problems previously discussed. The problem, therefore, is how can the papillary dermal thickness be measured noninvasively? In formulating a solution to this problem it is useful to recall that the amount of light remitted from the skin becomes highly dependent upon this factor as the wavelength increases [3]. When this is combined with the observation that both melanin and blood absorption drop signi cantly with increasing wavelength it seems it may be possible to nd a wavelength range where the amount of remitted light depends largely on the papillary dermal thickness. This is never the case within the visible portion of the spectrum. However, it becomes plausible if the considered wavelength range is extended into the infrared. For instance, in the wavelength range 600{800 nm the absorption of melanin drops to around one tenth of its peak within the visible portion of the spectrum; the absorption of blood drops by around a factor of a hundred whilst the sensitivity of remitted light to variations in papillary dermal thickness increases. This di erence becomes even more marked as longer wavelengths are considered; for example, the absorption of melanin drops by a further order of magnitude in the wavelength range 800{1000 nm. Indeed, within these wavelength ranges the thickness of the papillary dermis is the major parameter a ecting skin coloration. As these wavelength ranges are easily accessible with existing infrared lm and infrared digital cameras it should be possible to use this information to provide the desired calibration. As an example of how this may be performed consider Figure 1 where the intensity of remitted light for in vivo skin between 600{800 nm is plotted against that measured between 800{1000 nm. As can be seen from this graph the amount of remitted light falls, with increasing melanin, faster for the 600{800 nm primary than for the 800{1000 nm primary. This is as one would expect; more interesting, however, is the signi cant variation in both primaries with papillary dermal thickness. This observation then allows construction of the lines of constant papillary dermal thickness as shown in the graph. A measurement of this thickness parameter can thus be recovered by obtaining images acquired within the given wavelength ranges and looking-up the corresponding papillary dermal thickness from the graph.