INSTITUTE OF PHYSICS PUBLISHING Phys. Med. Biol. 50 (2005) N345–N357
PHYSICS IN MEDICINE AND BIOLOGY
doi:10.1088/0031-9155/50/22/N01
NOTE
Automated segmentation of pigmented skin lesions in multispectral imaging Mauro Carrara1, Stefano Tomatis1, Aldo Bono2, Cesare Bartoli3, Daniele Moglia2, Manuela Lualdi1, Ambrogio Colombo1, Mario Santinami2 and Renato Marchesini1 1
Medical Physics Unit, Istituto Nazionale per lo Studio e la Cura dei Tumori , Milan, Italy Melanoma and Sarcoma Unit, Istituto Nazionale per lo Studio e la Cura dei Tumori , Milan, Italy 3 Day surgery Unit, Istituto Nazionale per lo Studio e la Cura dei Tumori, Milan, Italy 2
E-mail:
[email protected]
Received 30 June 2005, in final form 31 August 2005 Published 2 November 2005 Online at stacks.iop.org/PMB/50/N345 Abstract The aim of this study was to develop an algorithm for the automatic segmentation of multispectral images of pigmented skin lesions. The study involved 1700 patients with 1856 cutaneous pigmented lesions, which were analysed in vivo by a novel spectrophotometric system, before excision. The system is able to acquire a set of 15 different multispectral images at equally spaced wavelengths between 483 and 951 nm. An original segmentation algorithm was developed and applied to the whole set of lesions and was able to automatically contour them all. The obtained lesion boundaries were shown to two expert clinicians, who, independently, rejected 54 of them. The 97.1% contour accuracy indicates that the developed algorithm could be a helpful and effective instrument for the automatic segmentation of skin pigmented lesions.
1. Introduction The incidence of malignant melanoma, a disease of worldwide distribution, has been rapidly increasing over the last few decades, especially among Caucasians (Jemal et al 2005). Early detection and prompt surgery represent the only curative management of patients affected by the disease (NIH Consensus Conference 1992, Balch et al 2001). The clinical diagnosis of melanoma is commonly based on the ABCD system, a checklist of parameters (asymmetry, border, colour and dimension) which was introduced to facilitate the visual recognition of this tumour (Friedman et al 1985). Improvements in diagnostic sensitivity have been reported with the application of dermatoscopy (Pehamberger et al 1993, Nachbar et al 1994, Cristofolini et al 1994). To evaluate whether a pigmented lesion is a melanoma or not is to a large extent subjective, as with any clinical diagnosis, and the experience of the individual clinician plays a fundamental role in the recognition of this disease. In the last decade, several techniques 0031-9155/05/220345+13$30.00 © 2005 IOP Publishing Ltd Printed in the UK
N345
N346
M Carrara et al
based on acquisition and elaboration of images of skin moles have been developed to obtain a 2D visualization of the lesions in an attempt to objectively evaluate the clinical features involved, such as those presented in Cascinelli et al (1992), Green et al (1994), Sober and Burstein (1994), Marchesini et al (1995), Bono et al (1996, 1999, 2002), Andreassi et al (1999), Binder et al (2000), Farina et al (2000), Rubegni et al (2001), Moncrieff et al (2002), Rosado et al (2003) and Tomatis et al (1998, 2003, 2005). In addition, for a cross-sectional representation of in vivo biological tissues, optical coherence tomography systems are under investigation. Details and/or preliminary results are presented in Welzel (2001), Marchesini et al (2002) and Marghoob et al (2003). A great variability exists in the visual appearance of pigmented skin lesions, ranging from very pale to highly dark colours, from homogeneously distributed to strongly variegated pigmentation, from well defined to soft vanishing lesion edges, from lesions smaller than a pin’s head to lesions with diameters larger than 15 mm. Therefore, the implementation of an accurate automatic segmentation method for skin pigmented lesions is a difficult task, and it is still a matter of research. In addition, clinicians do not always show a reliable intra- and inter-operator agreement in delineating lesion edge. Thus, an automatic approach is required able to perform reproducible outlines as well as reproducible extraction of lesion features, especially when infrared images are involved. Different segmentation techniques based on the intensity characteristics of the images may be applied (Gonzalez and Wood 1992), or have been proposed for skin lesion segmentation (Claridge et al 1992, Schindewolf et al 1993, Ercal et al 1993, 1994, Hall et al 1995, Elbaum et al 2001, Ganster et al 2001, Erkol et al 2005). Most of them refer to the grey-level histogram of the lesion image, evaluating gradients, peaks and local minima of these distributions. Limits of these techniques have also been discussed (Gonzalez and Wood 1992, Hall et al 1995). In this research, we evaluated the capability to correctly contour a skin lesion and its internal isles of dark colour by a different segmentation algorithm implemented in a novel spectrophotometric system. 2. Material and methods 2.1. Patients Between September 2002 and November 2004, 1700 patients (943 females and 757 males) with 1856 cutaneous pigmented lesions, including 264 melanomas, were enrolled in the study at the Istituto Nazionale Tumori of Milan. The median age of the patients was 36 years, ranging between 5 and 88 years. All the acquired lesions were consequently seen by the clinicians working in our Melanoma Unit. Inclusion criteria were based on clinical and/or dermatoscopic features that suggested a more or less important suspicion for melanoma. All the lesions were then subjected to surgery and histological examination. Lesions excluded from the study were those awkwardly situated, like those placed in the interdigital spaces, on ears, on the nose, in the navel, etc. If the lesion and the surrounding skin were hairy, to not interfere with reflectance measurement hairs were shaved with a razor before acquiring the image. The maximum diameter of the lesions ranged from 1 mm to 20 mm; mean and median diameters were 6.8 mm and 6 mm, respectively. Before surgery, images of the 1856 pigmented lesions were acquired in vivo. The slides were evaluated according to the widely accepted criteria for the histopathological diagnoses of the various pigmented lesions (Elder and Murphy 1991). Of the 264 melanomas 21 were in situ and 203 were thin lesions (Breslow thickness
