Elsevier

Pattern Recognition

Volume 64, April 2017, Pages 92-104
Pattern Recognition

A simple weighted thresholding method for the segmentation of pigmented skin lesions in macroscopic images

https://doi.org/10.1016/j.patcog.2016.10.031Get rights and content

Highlights

  • A simple and yet effective segmentation method for skin lesions is proposed.

  • A novel intensity map helps to discriminate lesions from background skin.

  • A multiple thresholding scheme helps to balance the contribution of the classes.

  • Segmentation relies on a modified version of Otsu's functional.

  • The proposed method may be used in clinical settings as part of CAD system.

Abstract

This work proposes a simple and yet effective thresholding method to segment pigmented skin lesions in macroscopic photographs automatically. Segmentation is one of the first steps in computer-aided diagnosis of skin cancers. Therefore, an accurate segmentation may play an important clinical role. We develop an algorithm that searches for a thin rectangular-shaped region near the image borders that is likely to contain mostly skin pixels. Segmentation is obtained by adapting Otsu's thresholding method by combining independent threshold estimates computed from histograms of different parts of a new intensity image designed to discriminate lesions from background skin. The proposed approach exploits the fact that the object of interest is approximately centered in the input photograph. A cross-diagonal sampling scheme helps to balance the size of the classes when the area of the lesion and the area of the surrounding skin are very different. A post-processing stage that includes morphological filtering and a weighted scheme to select the most salient object follows. The experimental results suggest that the method potentially can be used successfully to segment atypical nevi and melanomas in lesions with a highly heterogeneous background skin. The proposed algorithm is of interest for use in clinical settings as part of a CAD system.

Introduction

The World Health Organization estimates between 2 and 3 million non-melanoma skin cancers and 132,000 melanoma skin cancers occur globally each year. One in every three cancers diagnosed is a skin cancer, and the global incidence of melanoma continues to rise.1 Early detection is of paramount importance for melanoma treatment. At this stage, the prognosis for the patient is excellent as it can be cured by simple excision. However, melanoma detection in early stages is difficult because they resemble common nevi. Algorithms and computer-aided diagnosis systems (CADs) have been developed to support heath care professionals in this challenging task.

Dermoscopy, a non-invasive examination technique that uses a hand-held magnifier illuminated by a light source, allows doctors to see and evaluate colors and microstructures deeper in the skin lesion, not visible to naked eye. This technique has proven to increase diagnostic accuracy in hands of trained practitioners. Currently, most CADs systems use digital dermoscopy [1]. These CADs systems, however, have yet to demonstrate their benefit for experienced dermatologists in their clinical settings [2].

Diagnosis accuracy based on dermoscopy decreases in hands of non trained personal [3]. Therefore, having CADs that would be able to analyze macroscopic (clinical) digital photographs may complement dermoscopy-based systems and promote early melanoma detection. The use of standard digital cameras for acquiring images is appealing. Cheap and portable, they are increasingly accessible to the general population. A recent example is the use of cameras in smartphones to promote health applications. While experts remain cautious about the utility and safety of such “apps”, Kassianos et al. [4] noted that several of these have “the potential to take and store images of skin lesions, either for review by a dermatologist, or for self-monitoring to identify change, an important predictor of melanoma”. Some also provide risk assessment to patients, estimating the probability that a lesion was malignant or benign.

Image segmentation is a key component of CADs for the analysis of medical images [1], [5]. In case of skin lesions, segmentation amounts to finding the border that separates the lesion area from the surrounding skin. Once the lesion is detected, features describing the asymmetry, border, color, and structures present in the lesion can be computed and used to train statistical models to predict the diagnosis. Therefore, obtaining an accurate segmentation of the lesion is important, especially to quantify shape and border features [6].

Many algorithms have been proposed to segment pigmented skin lesions automatically. Morphological differences in the appearance of pigmented skin lesions in macroscopic and dermoscopic images directly influence the choice of method for lesion segmentation [1]. Otsu's thresholding [7] is a simple and fast approach and frequently has been used as a component in several algorithms designed for analysis of macroscopic photographs [8], [9], [10], [11], [12]. Thresholding methods determine a cutting value to turn a grayscale representation of an image into a binary image [13]. As an illustration, Ruiz et al. [8] applied Otsu's thresholding to grayscale images as part of a decision support system for the diagnosis of melanoma.

Alcón et al. [14] argue that thresholding does not suffer the drawbacks of edge detection-based algorithms that perform poorly when applied to skin images, mainly because of the presence of fine details, hair, or both on the skin. Snakes or active contours [15] are an alternative to thresholding to detect the complex shape of the suspicious melanoma lesions. However, the performance of snakes or active contours methods may be more sensitive to the presence of hair.

Considering the aforementioned observations, Cavalcanti and Scharcanski [11] proposed a method based on Otsu's thresholding applied to a color space projection that attempts to maximizes the separability between non-lesion and lesion pixels, followed by the Chan-Vese method. The algorithm was tested on the same data set used in this study. This approach was shown to provide segmentation results that were more accurate than previous methods, including a multi-direction gradient vector flow (GVF) snake-based scheme [16], a thresholding method inspired by Otsu's algorithm [14], [17] , thresholding a multichannel image representation to distinguish the skin lesion from the background areas [10], independent component analysis (ICA) followed by segmentation using the Chan-Vese method [9], and k-means on the image patches projected in a learned dictionary obtained via non-negative matrix factorization (NMF) [18].

Recently, Flores and Scharcanski [19] used a dictionary-based technique to represent a compact description of image patches. The authors proposed an unsupervised version of to the information-theoretic dictionary learning (ITDL) method [20], and segmented the skin lesions using normalized graph cuts [21].

In addition to color information, another interesting alternative is to incorporate textural information in the segmentation process [1]. For instance, Glaister et al. [12] used the Otsu's thresholding method combined with statistical region merging [22] to segment skin lesions in macroscopic images.

We propose a new automated algorithm for the segmentation of pigmented skin lesions in macroscopic photographs, namely Simple Weighted Otsu Thresholding (SWOT), which adapts the original Otsu's thresholding functional considering (i) all samples in the image and (ii) samples located in the cross-diagonals of the image. The resulting threshold is weighted with an independent estimate obtained from a peripheral region containing mainly skin pixels. While strikingly simple, SWOT is shown to yield segmentation accuracy in data set used by Alcón et al. [14] that are very similar to or better than previous methods. Competitive results also have been obtained on the Dermquest data set. Furthermore, it is fast and could operationally run on equipment with limited processing and memory resources. In addition, SWOT is easy to use and straightforward to extend to multispectral images.

The remaining of the paper is organized as follows. Section 2 reviews Otsu's thresholding method. Section 3 briefly reviews the preprocessing algorithm used to attenuate shadows in the input photographs. Section 4 presents the novel contributions of this work, which includes (i) a description of an automated procedure to select a thin region near the border of the image, that is likely to contain mostly skin pixels. Samples from this region are used to estimate the color of the background skin, used to compute the proposed intensity image, that is the input to the (ii) new thresholding algorithm. Section 5 describes the post-processing scheme used to eliminate possible artifacts, selecting the continuous object that is larger and closer to the center of the image. Section 6 investigates how the parameters setting affect the accuracy of the proposed segmentation, and how the accuracy compares with state-of-the-art algorithms. After discussing results in Section 7, we conclude in Section 8.

Section snippets

Threshold selection

Given the pixels of a grayscale image, represented by L gray levels {1,2,3,,L}, segmentation aims at splitting the pixels into two classes C1 and C2 (e.g. background and foreground). In this context, C1 denotes the pixels with levels {1,2,3,,t} and C2 denotes the pixels with levels {t+1,,L}.

Otsu's thresholding method is a popular method to solve such a problem. Working on the histogram of an image, Otsu's method exhaustively searches for the partition threshold t that minimizes the

Pre-processing: shadow attenuation

Illumination variations hamper the segmentation of skin lesions in macroscopic photographs. Therefore, the input images to be segmented using all methods tested in Section 6 were first corrected for shadow attenuation using the approach proposed in [25], which compensates for non-uniform illumination. The algorithm assumes that the Value component V(x,y) of the images in the HSV color space can be modeled by a quadratic formulation z(x,y), as a function of the spatial coordinates (x,y) of the

SWOT segmentation

The proposed SWOT segmentation is characterized by five processing steps:

  • 1.

    Automatic selection of candidate skin pixels.

  • 2.

    Computation of an intensity image for thresholding.

  • 3.

    Noise reduction using median filtering.

  • 4.

    Threshold estimation.

  • 5.

    Post-processing.

Fig. 1(a) shows a flowchart of the proposed SWOT segmentation method showing the main processing steps. The boxes in gray show the novel contributions of our study. 4.1 Automatic selection of candidate skin pixels, 4.2 Computation of an intensity image

Post-processing stage

The binary segmentation mask, produced applying the threshold t (Eq. (8)) to the intensity image I (Eq. (7)), might contain several disjoint objects that could be confused with the lesion region.4 The main objective of the post-processing stage is to eliminate small/thin objects, that might be remaining of hair or other artifacts, and select the single connected region that is more likely to be the actual lesion.

We start

Macroscopic images

Two data sets were used for experiments. The first was the data set used by Alcón et al., that consists of 152 macroscopic photographs of pigmented skin lesions, with benign and malignant diagnosis. Out of these, 45 images correspond to atypical nevi, and the remaining 107 were melanoma. Each 24 bits RGB color image has a size of 720 pixels on the larger side, and between 439 and 706 pixels on the smaller side. Fig. 2, Fig. 3 show two examples. Some images were duplicates from the same lesion,

Discussion

Ease of use is an important property of the proposed SWOT algorithm. It has two free parameters (α,β) that can be easily estimated using a quantitative or qualitative assessment on a separate data set. For convenience, all the remaining secondary parameters, like the filter sizes and the width of the cross-diagonal region area, are set in terms of the relative size of the image (we did not attempt to optimize such secondary parameters, and eventually better results could be obtained using

Conclusions

Segmentation is an important component in the computer assisted diagnosis of skin cancers. We proposed a new method for segmentation of macroscopic photographs of pigmented skin lesions, SWOT, which is based on a modification of the Otsu's thresholding method, analyzing separately samples from different locations in the image. SWOT provided good segmentation results for most of the 152 and 137 melanocytic lesions in the Alcón et al. and Dermquest test sets, respectively, including benign and

Acknowledgments

This work was supported in part by the Brazilian National Council for Scientific and Technological Development (CNPq) (grant number 401113/2014-0) .

Maciel Zortea is a researcher developing image analysis and pattern recognition methodologies applied in remote sensing and dermatology. He obtained his PhD from the University of Genoa, Italy, in 2007. He is currently with the Institute of informatics at the Federal University of Rio Grande do Sul, Porto Alegre, Brazil.

References (30)

  • W.G. Flores et al.

    Improving classification performance of breast lesions on ultrasonography

    Pattern Recognit.

    (2015)
  • N. Otsu

    A threshold selection method from gray-level histograms

    IEEE Trans. Syst. Man Cybern.

    (1979)
  • D. Ruiz, V. Berenguer, A. Soriano, J. Martin, A cooperative approach for the diagnosis of the melanoma, in: Proceedings...
  • P.G. Cavalcanti, J. Scharcanski, L. Di Persia, D. Milone, An ICA-based method for the segmentation of pigmented skin...
  • J. Glaister et al.

    Segmentation of skin lesions from digital images using joint statistical texture distinctiveness

    IEEE Trans. Biomed. Eng.

    (2014)
  • Cited by (60)

    • Decoupling multi-task causality for improved skin lesion segmentation and classification

      2023, Pattern Recognition
      Citation Excerpt :

      Skin lesion analysis usually consists of three main tasks, as shown in Fig. 1, including bounding box detection for lesion’s position, mask segmentation for specific contour and lesion classification for benign/malignant type, where the detection task is often skipped due to the lack of lesion bounding box annotations. Many methods have been proposed for skin lesion segmentation [3,4] or classification tasks [5,6]. Among these methods, the deep learning based methods perform better than those based on hand-crafted features such as shape [7] or texture [8], and become dominant in the computer aided diagnosis gradually [9,10].

    • Knowledge-aware deep framework for collaborative skin lesion segmentation and melanoma recognition

      2021, Pattern Recognition
      Citation Excerpt :

      A CAD system for melanoma analysis generally contains two crucial functions: lesion segmentation and melanoma recognition. Specifically, the task of segmentation aims to divide a dermoscopic image into the skin lesion parts and background parts, i.e., it is a pixel-wise classification process to generate more conceptual saliency information for melanoma analysis [5–9]. Meanwhile, melanoma recognition is an image-level classification task that aims to identify the skin lesion types, such as melanoma, seborrheic keratosis, and benign nevi [10–14].

    View all citing articles on Scopus

    Maciel Zortea is a researcher developing image analysis and pattern recognition methodologies applied in remote sensing and dermatology. He obtained his PhD from the University of Genoa, Italy, in 2007. He is currently with the Institute of informatics at the Federal University of Rio Grande do Sul, Porto Alegre, Brazil.

    Eliezer Flores received the B.Eng. degree in computer engineering from the Federal University of Pampa, Bagé, RS, Brazil and the M.Sc. degree in computer science from the Federal University of Rio Grande do Sul, Porto Alegre, RS, Brazil, in 2013 and 2015, respectively. He is an Assistant Professor with the Department of Telecommunications Engineering, Federal University of Pampa, Alegrete, Brazil. His current research interest is medical imaging applications.

    Jacob Scharcanski is a Professor in Computer Science at the Federal University of Rio Grande do Sul (UFRGS), Brasil. He holds a cross appointment with the Department of Electrical Engineering at UFRGS, and also is an Adjunct Professor with the Department of Systems Design Engineering, University of Waterloo, Ontario, Canada. He has authored and co-authored over 150 refereed journal and conference papers, books and book chapters on imaging and measurements. In addition to his academic publications, he has several technology transfers to the private sector. Presently, he serves as an Associate Editor for two journals, and has served on dozens of International Conference Committees. Professor Scharcanski is a licensed Professional Engineer, Senior Member of the IEEE, IEEE IMS Distinguished Lecturer (2015–2017). His areas of expertise are Image Processing, Pattern recognition, Imaging Measurements and their applications.

    View full text