PDE-based unsupervised repair of hair-occluded information in dermoscopy images of melanoma

Abstract

The repair of hair-occluded information is one of the key problems for the precise segmentation and analysis of the skin malignant melanoma image with hairs. Aimed at dermoscopy images of pigmented skin lesions, an unsupervised repair algorithm for the hair-occluded information is proposed in this paper. This algorithm includes three steps: first, the melanoma image with hairs are enhanced by morphologic closing-based top-hat operator and then segmented through statistic threshold; second, the hairs are extracted based on the elongate of connected region; third, the hair-occluded information is repaired by the PDE-based image inpainting. As a matter of fact, with the morphologic closing-based top-hat operator both strong and weak hairs can be enhanced simultaneously, and the elongate state of band-like connected region can be correctly described by the elongate function proposed in this paper so as to measure the hair effectively. Therefore, the unsupervised repair problem of the hair-occluded information can be resolved very well through combining the hair extracting with the image inpainting technology. The experiment results show that the repaired images can satisfy the requirement of medical diagnosis by the proposed algorithm and the segmentation veracity is effectively improved after repairing the hair-occluded information.

Introduction

Malignant melanoma (MM) has consistently had one of the most rapidly increasing incidence of all cancers, with 62,190 new cases and 7700 deaths estimated in the United States in 2006, and the incidence of MM increased 3–8% per year and doubled over the decade in China [1], [2], [3]. Early diagnosis is particularly important since melanoma can be cured with a simple excision if detected early [4], [5].

Dermoscopy is a non-invasive skin imaging technique that uses optical magnification and either liquid immersion and low angle-of-incidence lighting or cross-polarized lighting to make the contact area translucent, making subsurface structures more easily visible when compared to conventional macroscopic (clinical) images [2]. Dermoscopy allows the identification of dozens of morphological features such as pigment networks, dots/globules, streaks, blue-white areas, and blotches [2]. But due to the lack of reproducibility and subjectivity of human interpretation, the development of computerized image analysis techniques is of paramount importance.

Digital image analysis technology about skin melanoma is currently focusing on segmentation, feature extraction and classifier design. Grana et al. [6] used Otsu's threshold to automatically segment the melanoma image, then selected k points for spline interpolating to get smooth lesion border, finally produced statistical parameters about lesion based on ABCD rules. Taouil et al. [7] used top-hat operator as image preprocessor to eliminate noise and preserve lesion border information, then used Snake method to detect lesion border, which got better result than Otsu's threshold. Schmid [8] proposed a technique based color clustering. First, a 2D histogram is calculated from the first two principal components of the CIE L*u*v* color space. Then, the histogram is smoothed and initial cluster centers are determined from the peaks using a perceptron classifier. Finally, the lesion image is segmented using a modified version of the fuzzy c-means (FCM) clustering algorithm. On the classifying aspect, neural network diagnosis of skin tumors has previously been applied by classifying extracted features from digitized dermoscopy images of lesions [9]. The extracted features are based on geometry, colors, and texture of the lesions, involving complex image processing techniques. Kusumoputro and Ariyanto [10] used an artificial neural network to separate the malignant melanoma from benign categories of skin cancers based on cancer shapes and their relative color. Celebi et al. [2], [11] used the Statistical Region Merging (SRM) algorithm to detect the lesion border, and realized the classification of dermoscopy images based on Support vector machine using the shape, color and texture as input features. And he pointed out that the hairs in the image will cause to a bad segmentation result so as to error analysis.

The hair problem has not been fully addressed in the literature of computer image processing techniques for dermatological applications in spite of the rapid growth of this field [12]. It is an important issue, however, especially for designing automatic segmentation and feature extraction algorithms for clinical use. In a similar study conducted in Italy [13], the investigators decided to shave the hairs using a razor before the imaging session. This procedure not only adds extra costs and time to the imaging session, but it also is impractical if we want to apply the technology to total-body nevus imaging [14]. Lee et al. [15] located the hair through morphological closing operator first and then removed the hairs from image by replacing the hair pixels by the nearby non-hair pixels, but the algorithm worked well just for the thick and dark hairs.

In this paper a novel unsupervised hair removal algorithm is proposed through combining hair object extraction with image inpainting technology. Image inpainting is a kind of effectively method for both repairing the lost information or nicks in image and removing some specified object from image. Strictly speaking, current image inpainting methods can hardly implement automatic processing performance because some area information needs to be inputted by user in order to determine which area is to be repaired [16], [17]. In this paper the hairs are extracted based on elongate feature function first and then the information of occluded area is successfully repaired through the image inpainting technology. The experiment results show that the extracted hairs are exact and the repaired texture satisfies the requirement of medical diagnosis.