A narrow band graph partitioning method for skin lesion segmentation

doi:10.1016/j.patcog.2008.09.006

Pattern Recognition

Volume 42, Issue 6, June 2009, Pages 1017-1028

https://doi.org/10.1016/j.patcog.2008.09.006 Get rights and content

Abstract

Accurate skin lesion segmentation is critical for automated early skin cancer detection and diagnosis. In this paper, we present a novel multi-modal skin lesion segmentation method based on region fusion and narrow band energy graph partitioning. The proposed method can handle challenging characteristics of skin lesions, such as topological changes, weak or false edges, and asymmetry. Extensive testing demonstrated that in this method complex contours are detected correctly while topological changes of evolving curves are managed naturally. The accuracy of the method was quantified using a lesion similarity measure and lesion segmentation error ratio, Our results were validated using a large set of epiluminescence microscopy (ELM) images acquired using cross-polarization ELM and side-transillumination ELM. Our findings demonstrate that the new method can achieve improved robustness and better overall performance compared to other state-of-the-art segmentation methods.

Introduction

Early detection of a cancerous skin lesion is crucial for successful treatment and cure. For example, malignant melanoma, the deadliest form of all skin cancers, has a cure rate of more than 95% if detected at an early stage [1]. To facilitate early detection, several methods have been developed for melanoma detection [2], [3], [4], [5]. In all these methods, delineating the skin lesion correctly is key to lesion analysis and accurate diagnosis.

Extracting the lesion area from the background is an essential step in all computer-aided melanoma detection systems and it has attracted continuous research efforts. Thresholding [6], [7], [8], [9], [10] and region growing are two simple yet widely used algorithms in the literature. They produce a satisfactory segmentation when skin lesions have clear boundaries. To account for noise and unclear boundaries, several clustering-based methods have been developed that demonstrated improved robustness [10], [11], [12]. Another family of popular segmentation methods is based on active contours (or snakes) [13], [14], [15], whereby a curve, defined based on partial differential PDEs evolves toward the local optimum, with respect to an objective function. Typical objective functions are chosen to identify the region of interest and are mostly edge-based or region-based. Chung [16] and Erkol [17] applied edge-based active contours to segment skin lesions. However, leakage occurs at the presence of weak edges and the method is sensitive to the initial conditions. Chan and Vese [18] proposed a region-based active contour method that models non-overlapping homogeneous regions with Gaussian distributions. The method performs well on images with two regions of distinct intensities; however, it presents under-segmentation problems if there are more than two regions of interest in the image. A hierarchical approach has also been proposed [19] to tackle this problem by splitting manually identified subregions.

Following our previous work on region-based object identification [20], [21], [22], we developed an active contour-based region-fusion framework. In this approach, a lesion is first segmented into small regions by imposing strict constraints on homogeneity and strong edges on the region-based snakes. Then, these regions are merged based on a centroid distance criterion and gradient information. To improve the performance and computational efficiency of segmentation, a narrow-band graph-partitioning (NBGP) method has been developed.

Experiments and comparison results on more than 200 skin lesion images demonstrate that the proposed algorithm produces segmentation results that are very close to the manual segmentation provided by specialists, even for those images with highly asymmetric lesions, weak and/or false edges, or strong hair and bubble artifacts.

The rest of the paper is organized as follows: in Section 2, we present a short review on skin lesion characteristics and image acquisition. In Section 3, we start with an introduction on the level set formulation of the region-based active contours, and then the derivation and development of the NBGP curve evolution for image segmentation is presented in detail. Section 4 describes the new region-fusion algorithm for skin lesion segmentation based on NBGP, whereas Section 5 presents experimental results on more than 200 images from three different imaging modalities. The results obtained with the new method are compared with results from our previously developed methods [10] and from active contours [18], [28] and validated against manual segmentation by expert dermatologists. Finally, Section 6 concludes this paper with a discussion on our findings and a summary of future work.

Section snippets

Characteristics of skin lesion images

In epiluminescence microscopy (ELM), a halogen light is projected into the object rendering the surface translucent [2]. There are two modes of ELM used in clinical applications, oil immersion [23], [24] and cross-polarization (XLM) [25]—the latter was developed to reduce light reflection from the epidermis. In side-transillumination ELM (TLM), a bright ring of light positioned around the periphery of a lesion is projected onto the center of the lesion at $45^{\circ}$ forming a virtual light source at a

Level set formulation of region-based curve evolution

Given an image $I \subset Ω$ , the region-based active contour model [18] aims to find the curve $\overset{\Rightarrow}{C}$ that minimizes the energy-based objective function: $E (\overset{\Rightarrow}{C}) = λ_{1} \iint_{inside} | I (x, y) - c_{1} |^{2} d x d y + λ_{2} \iint_{outside} | I (x, y) - c_{2} |^{2} d x d y + μ L (\overset{\Rightarrow}{C}) + v A (inside (\overset{\Rightarrow}{C}),$ where $c_{1}$ is the average intensity inside $\overset{\Rightarrow}{C}$ ; $c_{2}$ is the average intensity outside $\overset{\Rightarrow}{C}$ ; $μ$ , $v ⩾ 0$ ; and $λ_{1}$ , $λ_{2} > 0$ are fixed weight defined based on a priori knowledge. $L (\overset{\Rightarrow}{C})$ and $A (\overset{\Rightarrow}{C})$ are two regulation terms. Following Chan's approach [18], we can fix the weights as $λ_{1} = λ_{2} = 1$ , $μ > 0$ , and $v =$

Skin lesion segmentation algorithm based on region fusion of narrow band graph partitioning

In this section, we describe in detail a two-stage framework that integrates region fusion and narrow band graph partitioning (NBGP) active contour. The resulting skin lesion segmentation algorithm can efficiently and effectively segment challenging images as outlined in Section 2.

Image acquisition

Our XLM and TLM images were acquired with a Nevoscope [35], [36] that uses a $5 \times$ optical lens (manufactured by Nikon, Japan). An Olympus C2500 digital camera was attached to this Nevoscope. Fifty one XLM images and 60 TLM images were used in our experiments. The image resolution is $1712 \times 1368$ pixels. A region of interests (ROI), identified in every image, was segmented, manually by dermatologists and automatically using our method. This step was performed on a $236 \times 236$ region obtained from a

Discussion

In our experiments, we observed that active contours were attracted by structures on healthy skin or noise near the boundary. This is because the band size (ebs in Algorithm 2) used to compute the energy function is very small (since we chose a small ebs for the level set function update to reduce computational cost) compared to the dimension of the image, and thus the method is almost an edge based scheme. We mitigated the effect of edge leaking by using the multiscale approach.

Our method

Acknolwedgments

The authors are grateful to Dr. William Stoeker, M.D. and Nizar Mullani for providing the ELM, XLM and TLM images and clinical expertise.

References (42)

M. Elbaum et al.
Automatic differentiation of melanoma from melanocytic nevi with multispectral digital dermoscopy: a feasibility study
J. Am. Acad. Dermatol.
(2001)
X. Yuan et al.
Multi-scale feature identification using evolution strategies
Image Vision Comput. J.
(2005)
X. Yuan et al.
Evolution strategies based image registration via feature matching
Information Fusion
(2004)
H. Pehamberger et al.
In vivo epiluminescence microscopy of pigmented skin lesions. I. Pattern analysis of pigmented skin lesions
J. Am. Acad. Dermatology
(1987)
S. Osher et al.
Fronts propagating with curvature dependent speed: algorithms based on Hamilton Jacobi formulations
J. Comput. Phys.
(1988)
D. Adalsteinsson et al.
A fast level set method for propagating interfaces
J. Comput. Phys.
(1995)
T. Lee et al.
Dullrazor: a software approach to hair removal from images
Comput. Biol. Med.
(1997)
Anonymous, Cancer Facts and Figures 2002,...
H. Ganster et al.
Automated melanoma recognition
IEEE Trans. Med. Imaging
(2001)
S. Menzies et al.
The performance of solarscan: an automated dermoscopy image analysis instrument for the diagnosis of primary melanoma
Arch. Dermatol.
(2005)

M. Moncrieff et al.

Spectrophotometric intracutaneous analyses: a new technique for imaging pigmented skin lesions

Br. J. Dermatol.

(2002)

G. Betta, G. Leo, G. Fabbrocini, A. Paolillo, M. Scalvenzi, Automatic application of the ‘7-point checklist’ diagnosis...

R. Jailani, H. Hashim, M.N. Taib, S. Sulaiman, Border segmentation on digitized psoriasis skin lesion images, in: 2004...

E. Maglogiannis, C. Caroni, S. Pavlopoilos, V. Karioti, Utilizing artificial intelligence for the characterization of...

K. Taouil, N.B. Romdhane, Automatic segmentation and classification of skin lesion images,, in: The 2nd International...

G. Zouridakis, M. Doshi, N. Mullani, Early diagnosis of skin cancer based on segmentation and measurement of...

M.E. Celebi, Y.A. Aslandogan, P.R. Bergstresser, Unsupervised border detection of skin lesion images, in: ITCC (2),...

M.E. Celebi, Y.A. Aslandogan, P.R. Bergstresser, Mining biomedical images with density-based clustering, in: ITCC (1),...

M. Kass et al.

Snakes: active contour models

Int. J. Comput. Vision

(1988)

C. Xu et al.

Snakes, shapes, and gradient vector flow

IEEE Trans. Image Process.

(1998)

V. Caselles et al.

Geodesic active contours

Cited by (74)

Cascaded context enhancement network for automatic skin lesion segmentation
2022, Expert Systems with Applications
Skin lesion segmentation is an important step for automatic melanoma diagnosis. Due to the non-negligible diversity of lesions from different patients, extracting powerful context for fine-grained semantic segmentation is still challenging today. Although the deep convolutional neural networks (CNNs) have made significant improvements on skin lesion segmentation, they often fail to reserve the spatial details and long-range dependencies context due to consecutive convolution striding and pooling operations inside CNNs. In this paper, we formulate a cascaded context enhancement neural network for automatic skin lesion segmentation. A new cascaded context aggregation (CCA) module with a gate-based information integration approach is proposed to sequentially and selectively aggregate original image and multi-level features from the encoder sub-network. The generated context is further utilized to guide discriminative features extraction by the designed context-guided local affinity (CGL) module. Furthermore, an auxiliary loss is added to the CCA module for refining the prediction. In our work, we evaluate our approach on four public skin dermoscopy image datasets. The proposed method achieves the Jaccard Index (JA) of 87.1%, 80.3%, 84.3%, and 86.6% on ISIC-2016, ISIC-2017, ISIC-2018, and PH2 datasets, which show highly competitive performance with other state-of-the-art models respectively.
NURBS functional network approach for automatic image segmentation of macroscopic medical images in melanoma detection
2021, Journal of Computational Science
Image processing techniques are becoming standard technology in many medical specialities, such as dermatology, where they are a key tool for the early detection and diagnosis of melanoma and other skin cancers and tumors. A previous paper by the authors presented at SOCO 2020 conference introduced a new method for image segmentation of skin images through functional networks. The method performs well but it relies on a semi-automatic approach involving a combination of manual and automatic operations. This paper aims at making image segmentation of macroscopic skin images a fully automatic process. To this purpose, the present work extends our previous paper with five new relevant contributions: (1) a filtering strategy for removal of noise, hair and other artifacts; (2) two morphological operators for image enhancement; (3) a clustering-based binary classifier to separate the skin tumor from the image background; (4) a smoothing and discretization process to obtain the border points; and (5) a curve reconstruction method from the border points with NURBS using a new type of functional network particularly tailored for this task. This new method is applied to two different benchmarks, comprised respectively of four and two macroscopic medical images of skin tumors. The visual and numerical results show that the method performs very well, yielding segmented images which are suitable for clinical practice. This method is a significant step towards the future development of a fully automatic approach for the whole medical image analysis pipeline of skin images, including diagnosis and classification.
Particle method for segmentation of breast tumors in ultrasound images
2020, Mathematics and Computers in Simulation
Citation Excerpt :
The method eliminates the need for reinitialization of the contour required by the preceding methods and thereby avoids initialization-induced numerical errors. The recent modifications are the Neuro LSM [15], Narrow Band Graph Partitioning LSM [78], and Local Regional Fitting LSM [43], LSM preserving the distance function [20] and LSM employing a speed function based on saturation and lightness [48] (skin cancer images). The controlling parameters of the LSM estimated by a classic segmentation method, create hybrid schemes.
We propose a new segmentation method based on multiple walking particles (WP) bouncing from the image edges. The particles are able to segment objects characterized by deep concavities as narrow as one pixel and handle single or multiple objects characterized by a noisy background and broken boundaries (“weak edge”, “boundary leakage”). The particles are designed to segment the image by permanently staying inside the object and repairing the boundaries where necessary. The proposed WP combine the advantages of the continuous diffusion models with the principles of multi-agent systems.
WP have been tested against recent active contours and the distance regularized level set method on a set of complex-shaped synthetic images and ultrasound (US) images of breast cancer (http://onlinemedicalimages.com). The method has also been compared with localizing region-based active contours, the fuzzy C-mean level set method, and morphological active contours. The WP are faster and more accurate for images characterized by low contrast, noise, broken boundaries, or boundary leakage. However, for good quality, simple shaped objects the WP work similarly to the conventional methods. There is still an important difference even in this case: the WP do not require initialization.
A video demo of the algorithm is at https://drive.google.com/drive/folders/1SlTphINKtdUwvdjjxiakFrwI2dDlIUAU
Neutrosophic set in medical image clustering
2019, Neutrosophic Set in Medical Image Analysis
Neutrosophic sets (NS) have a significant role for denoising, clustering, segmentation, and classification in numerous medical image-processing applications. For efficient development of computer-aided diagnosis systems, clustering techniques have been incorporated with NS to diminish uncertainty for efficient diagnosis. To integrate the clustering methods with NS for further clustering, the image is transformed into the NS domain, which is defined by three membership functions (MFs): true (T), indeterminacy (I), and false (F). These MFs map the medical image into the NS domain to generate an NS image. During this process, different types of filters can be applied, such as the median filter to compute T, which iteratively determines the best version of the image of T and F functions. In addition, another filter can be used to compute the indeterminacy neutrosophic subset I, such as Sobel, Prewitt, and the unsharp filter. Such filters have a significant role in the determination of MFs. In the clustering stage, the pixels whose I values are greater than a specific threshold will be excluded from the clustering process. Other pixels are clustered using the K-means. Additionally, this chapter studies the effect of the filter type and size in NS during the calculation of I MF (uncertainty MF) and the segmentation results on the skin lesion regions in dermoscopic images using the NS-based K-means (NKM) clustering. A comparative study was conducted using these different types and/or sizes of these NS-based filters on the K-means clustering process on the ISIC 2016 skin lesion dermoscopic image dataset. The results established that the characteristics of the filter in the calculations of the indeterminacy neutrosophic subset I have a significant effect on the NS performance and the clustering process as well. Additionally, the results proved the superiority of the unsharp filter with the size of 7 × 7 compared to the other edge detection filters with an accuracy of 95.25% accuracy.
Step-wise integration of deep class-specific learning for dermoscopic image segmentation
2019, Pattern Recognition
The segmentation of abnormal regions on dermoscopic images is an important step for automated computer aided diagnosis (CAD) of skin lesions. Recent methods based on fully convolutional networks (FCN) have been very successful for dermoscopic image segmentation. However, they tend to overfit to the visual characteristics that are present in the dominant non-melanoma studies and therefore, perform poorly on the complex visual characteristics exhibited by melanoma studies, which usually consists of fuzzy boundaries and heterogeneous textures. In this paper, we propose a new method for automated skin lesion segmentation that overcomes these limitations via a novel deep class-specific learning approach which learns the important visual characteristics of the skin lesions of each individual class (melanoma vs. non-melanoma) on an individual basis. We also introduce a new probability-based, step-wise integration to combine complementary segmentation results derived from individual class-specific learning models. We achieved an average Dice coefficient of 85.66% on the ISBI 2017 Skin Lesion Challenge (SLC), 91.77% on the ISBI 2016 SLC and 92.10% on the PH2 datasets with corresponding Jaccard indices of 77.73%, 85.92% and 85.90%, respectively, for the same datasets. Our experiments on three well-established public benchmark datasets demonstrate that our method is more effective than other state-of-the-art methods for skin lesion segmentation.
A survey on automated melanoma detection
2018, Engineering Applications of Artificial Intelligence
Citation Excerpt :
Various methods have been proposed for lesion segmentation, which is principally an image segmentation task. Typical examples are based on thresholding (Garnavi et al., 2011; Celebi et al., 2013), clustering (Schmid, 1999; Mete et al., 2011; Melli et al., 2006), fuzzy logic (Baral et al., 2014), supervised learning (Wu et al., 2013), and graph theory (Yuan et al., 2009). In addition, some of these methods can be combined to achieve higher segmentation accuracy (Celebi et al., 2009a).
Skin cancer is defined as the rapid growth of skin cells due to DNA damage that cannot be repaired. Melanoma is one of the deadliest types of skin cancer, which originates from melanocytes. While other skin cancer types have limited spreading capabilities, the danger of melanoma comes from its ability to spread (metastasize) rapidly. Fortunately, it can be detected by visual inspection of the skin surface, and it is 100% curable if identified in the early stages. However, detection by “subjective visual inspection” creates an important problem, due to investigators’ different levels of experiences and education. Dermoscopy (dermatoscopy) has significantly increased the diagnostic accuracy of melanoma since late 90’s. In addition, several systems have been proposed in order to assist investigators or to perform an automatic melanoma detection. This survey focuses on the algorithms for automated melanoma detection in dermoscopic images through an extensive analysis of the stages in methodologies proposed in the literature, and by examining related concepts and describing possible future directions through open problems in this domain of research.

View all citing articles on Scopus

About the Author—XIAOJING YUAN received B.S. in electrical engineering from Hefei University of Technology in China in 1994, followed by a M.S. in computer engineering from University of Science and Technology of China in 1997 and another M.S. degree computer science in 2001 and a Ph.D. degree in Intelligent Systems in 2003, both from Tulane University. She joined the Engineering Technology Department of the University of Houston in 2004, where she has been actively involved in research areas such as biomedical imaging analysis, data mining, pattern recognition, and information technology in biomedical applications. She founded and served as director for “Intelligent Sensor Grid and Informatics Lab” in 2005. She has years’ experience in embedding intelligence into sensors and actuators to deal with uncertainties. She authored and co-authored more than 40 technical papers; has one patent and another one filed in 2008. She has been very active in professional organization such as IEEE and ISA; being reviewer for top-rank journals and conferences, and chairing and co-chairing conference sessions in “intelligent sensor network”.

About the Author—NING SITU received a B.S. degree in Computer Science from Sun Yat-sen University in 2006. He has been a Ph.D. student at the Computer Science Department, University of Houston since 2006. His current research interests include pattern recognition, image segmentation, and statistical imaging analysis for cancer detection.

About the Author—GEORGE ZOURIDAKIS received a Dr.Ing. degree in Electronics Engineering from the University of Rome “La Sapienza” in 1987, followed by an M.S. in Biomedical Engineering in 1990 and a Ph.D. in Electrical Engineering in 1994, both from the University of Houston. He has been on the faculty of The University of Texas-Houston Medical School from 1994 to 2001, where his clinical activities included Intraoperative Monitoring, Functional Brain Mapping, and Deep Brain Stimulation. He joined the University of Houston in 2001 and he is the Director of the Biomedical Imaging Lab. His current research interests are in the areas of Biomedical Imaging, Computational Biomedicine, Functional Brain Mapping, and Biosignal Analysis and Modeling. He is the main author of a book on Intraoperative Monitoring, CRC Press, 2001, and the Co-Editor-in-Chief of the “Handbook of Biomedical Technology and Devices”, CRC Press, 2003. He has developed courses, given lectures, organized sessions at national and international conferences on Medical Imaging and Brain Mapping, and has published more than 160 referred papers and abstracts. He is an Associate Editor of the IEEE Transactions on Biomedical Engineering, the Chair of the IEEE EMBS Houston Chapter, and he is also listed on Who's Who in America.

^☆: This work has been supported in part by NSF Grant 521527, the Grants to Enhance and Advance Research program, and the Texas Learning and Computation Center at the University of Houston.

View full text

Published by Elsevier B.V.

A narrow band graph partitioning method for skin lesion segmentation☆

Abstract

Introduction

Section snippets

Characteristics of skin lesion images

Level set formulation of region-based curve evolution

Skin lesion segmentation algorithm based on region fusion of narrow band graph partitioning

Image acquisition

Discussion

Acknolwedgments

J. Am. Acad. Dermatol.

Image Vision Comput. J.

Information Fusion

J. Am. Acad. Dermatology

J. Comput. Phys.

J. Comput. Phys.

Comput. Biol. Med.

Automated melanoma recognition

IEEE Trans. Med. Imaging

The performance of solarscan: an automated dermoscopy image analysis instrument for the diagnosis of primary melanoma

Arch. Dermatol.