Automated prescreening of pigmented skin lesions using standard cameras

Abstract

This paper describes a new method for classifying pigmented skin lesions as benign or malignant. The skin lesion images are acquired with standard cameras, and our method can be used in telemedicine by non-specialists. Each acquired image undergoes a sequence of processing steps, namely: (1) preprocessing, where shading effects are attenuated; (2) segmentation, where a 3-channel image representation is generated and later used to distinguish between lesion and healthy skin areas; (3) feature extraction, where a quantitative representation for the lesion area is generated; and (4) lesion classification, producing an estimate if the lesion is benign or malignant (melanoma). Our method was tested on two publicly available datasets of pigmented skin lesion images. The preliminary experimental results are promising, and suggest that our method can achieve a classification accuracy of 96.71%, which is significantly better than the accuracy of comparable methods available in the literature.

Introduction

Pigmented skin lesions include both, benign and malignant forms. Melanoma is a kind of malignant pigmented skin lesion, and currently is among the most dangerous existing cancers, resulting in about 10,000 deaths from the 40,000 to 50,000 diagnosed cases per year, just in United States of America [1]. According to World Health Organization [2], about 132,000 melanoma cases occur globally each year. Benign pigmented skin lesions are called moles, or nevi. However, differentiating benign and malignant lesions can be challenging. For example, there are nevi known as Clark Nevi (also referred as Dysplastic or Atypical Nevi) that present similar characteristics to melanomas [3]. It is consensual that the early diagnosis of malignant skin lesions (melanomas) is essential for the patient prognosis.

Recently, telemedicine techniques have been studied as a resource to obtain an early diagnosis of skin lesions. Besides the fact that dermatology probably is the most visual specialty in medicine, the teledermatology consultation brings some benefits, like easier access to health care and faster clinical results [4]. Comparing the physical examination (face-to-face diagnosis) with the remote diagnosis, experiments indicate that teledermatology is effective and reliable [5]. Teledermatology can benefit from image prescreening to help identify potentially malignant cases in their early stages.

To help distinguishing between benign and malignant cases, dermatologists often analyze each lesion with a dermoscope, which is a noninvasive tool that facilitates the evaluation of submacroscopic morphologic and vascular structures. As can be seen in Fig. 1, dermoscopy enables the generation of images with constant illumination, different texture patterns, and characteristics that are not measurable in standard camera images, such as lesion area and perimeter. In this way, many research groups developed digital dermoscopy image analysis schemes to help in skin lesion diagnosis [6].

In an attempt to prescreen/classify dermoscopy images, Celebi et al. [8] achieved 92.34% and 93.33% of specificity and sensitivity, respectively, using a JSEG-based segmentation algorithm and Support Vector Machines in the classification. More recently, Iyatomi et al. [9] proposed, to the best of our knowledge, the first publicly accessible system (“Dermatologist-like”), based on a region growing segmentation method and an Artificial Neural Network classifier. The user can upload an image at their website ‘http://dermoscopy.k.hosei.ac.jp’ and obtain a prescreening result. This system achieved a sensitivity of 85.9% and a specificity of 86.0% for a set of 1258 dermoscopy images [9].

Despite the importance of these efforts, a disadvantage of these methods is that they require dermoscopy images, and dermoscopes are not common among non-specialists. Moreover, studies indicate that dermoscopy images do not increase diagnosis accuracy in early stages [10]. So, with the proposal to facilitate the access to health care, also have been developed teledermatology systems making use of images acquired with standard cameras. In this way, patients do not need to go physically to a hospital or a clinic for a preliminary evaluation (even in benign cases), specially in remote areas. However, due to the already mentioned different visible characteristics in standard camera images and in dermoscopy images (see Fig. 1), these systems require different segmentation methods and feature extraction techniques.

A recent approach proposed by Alcon et al. [11] is an easy-to-use melanoma prescreening system based on standard camera images. A skin lesion photograph is provided as an input, and its prescreening is automatically produced, using segmentation and classification algorithms. However, often the acquired images contain artifacts, such as uneven illumination, which causes difficulties in the lesion segmentation stage, therefore their system initially corrects the image background. Afterwards, 55 features are extracted and the ABCD rule (Asymmetry, Border irregularity, Color variation and Differential structures) is employed to classify the lesion image as benign or malignant. Their system reaches an accuracy of 86.64% in its best performance [11].

This paper describes a new melanoma prescreening method using standard camera images and new techniques to improve the processing and analysis of such images, which was designed to be used remotely by non-specialists. In our experiments (see Section 6), we used 220 images obtained from two websites, with no special care in image acquisition or postprocessing. Each one of these images was submitted to a sequence of processing steps, namely: (1) preprocessing, where shading effects are attenuated by a new preprocessing stage proposed in this paper, as described in Section 2; (2) segmentation, where a new 3-channel image representation is proposed and used to discriminate between lesion and healthy skin areas, as described in Section 3; (3) feature extraction, where a quantitative lesion description containing new features extracted from our 3-channel representation is generated, as described in Section 4; and finally (4) lesion classification provides a lesion pre-diagnosis (using a hybrid classifier proposed in this paper); this processing step was designed to reduce the number of false negatives in the classification of skin lesion images as malignant or non-malignant, as described in Section 5. In Section 6, experimental results are presented, and Section 7 we present our conclusions.