Abstractâ Diagnosing skin cancer in its early stages is a challenging task for dermatologists given the fact that the chance for a patient's survival is higher and ...
Pigment network-based skin cancer detection Naser Alfed, Fouad Khelifi, Member, IEEE, Ahmed Bouridane, Senior Member, IEEE, and Huseyin Seker, Member, IEEE
Abstract— Diagnosing skin cancer in its early stages is a challenging task for dermatologists given the fact that the chance for a patient’s survival is higher and hence the process of analyzing skin images and making decisions should be time efficient. Therefore, diagnosing the disease using automated and computerized systems has nowadays become essential. This paper proposes an efficient system for skin cancer detection on dermoscopic images. It has been shown that the statistical characteristics of the pigment network, extracted from the dermoscopic image, could be used as efficient discriminating features for cancer detection. The proposed system has been assessed on a dataset of 200 dermoscopic images of the ‘Hospital Pedro Hispano’ [1] and the results of cross-validation have shown high detection accuracy.
features for classifying melanoma skin cancer images and normal images. The next section describes the proposed system along with the directional Gabor filters which are used at the pre-processing stage as well as at the pigment network extraction stage. Section III reports the experimental results and section IV concludes the paper. II. PROPOSED SYSTEM The system proposed in this work consists of four main stages (see Fig. 1): 1) pre-processing 2) Pigment network extraction 3) Feature extraction and 4) Classification.
Original Image
Index Terms-Pigment network detection, dermoscopic image.
Processed Pre-processing Image
Pigment Network
No No
I. INTRODUCTION
Image
Extracted Yes Yes
Skin cancer is one of the most common forms of cancer, and may take the form of benign or melanoma tumours. The benign type can be considered less dangerous than malignant melanoma and can be cured successfully, whereas malignant melanoma is the deadliest form of skin lesion[2], [3]. In Australia, skin cancer has one of the highest mortality rates compared with other cancer types [4]. More than 221,000 people were diagnosed with cancer in 2005 and about 12,513 people died in New South Wales, Australia [5]. In 2010, around 12,800 cases of skin cancer were diagnosed in the UK [6]. In the United States, about 60,000 new cases were found with an estimated 8110 deaths in 2007 [7]. These numbers had increased by 2013 reaching a peak of 76,690 people diagnosed of whom 9480 people died [8]. Also, the time, cost and effort required for dermatologists to check all patients for skin cancer in this way are prohibitive. The automated detection of skin lesions has shown accurate results and helps dermatologists in better decision making[9],[10],[11]. Recently, there has been a dramatic increase in the availability and wide spread use of new technologies such as the use of Computer Aided Diagnosis systems which leads researchers to try diagnosing tumours; because they can be used by dermatologists to support their decision and are less time-consuming[12]. The pigment network can be considered as one of the crucial importance dermoscopic structures [13]. In this paper we propose an automated skin cancer detection system on Dermoscopic images using the pigment network structure which is extracted via directional Gabor filters[13]. It has been found out that the statistical characteristics of pigmented networks obtained from the processed images can be used as efficient discriminating
978-1-4244-9270-1/15/$31.00 ©2015 IEEE
Pigment Network Extraction
Decesion
Feature Extraction
Classification Features
Figure 1. Block diagram of the system
1) Pre-processing The first stage of the proposed system aims to enhance the original medical image by removing air bubbles, artifacts caused by applied gel before capturing images in addition to hairs, and other noise. Pre-processing is very important to avoid affecting the extraction of the pigment network by reflection artifacts. The blue component is the best plane in RGB representation that can be used to achieve better results from dermoscopy images [14]. Barata and et al. proposed two steps for reflective artifacts and hair detection [13]. The same techniques for reflection artifacts, hair detection and pigment network extraction have been adopted in this work. a) Reflective artifacts detection and removal A simple thresholding algorithm for artifacts reflection detection and removal is used. Every single pixel (x, y) can be detected and classified as a reflection artifact according to the following condition: {𝐼(𝑥, 𝑦 > 𝑇𝑅1 )} 𝑎𝑛𝑑 {(𝐼(𝑥, 𝑦) − 𝐼𝑎𝑣𝑔 (𝑥, 𝑦)) > 𝑇𝑅2 } Where I is the image, 𝐼𝑎𝑣𝑔 (𝑥, 𝑦) is the average intensity of the selected pixel’s neighborhood that is computed using a local mean filter with dimensions 11×11 and 𝑇𝑅1 = 0.7 , 𝑇𝑅2 = 0.098 are threshold values which have been obtained from [13]. Once the locations of artifacts are
7214
detected, an inpainting operation is applied accordingly to the information of their detected pixels’ neighbourhood, avoiding the influence of the artifacts. Fig. 2. Illustrates the process of converting the original colour image (a) into a blue component image (b) and then removal of artifacts were obtained in (c).
(a)
(b)
(4) 𝐼𝑖 (𝑥, 𝑦) = ℎ𝜃𝑖 (𝑥, 𝑦) ∗ 𝐼(𝑥, 𝑦) After filtering the blue component image I, the output J which is a combination of the output of N + 1 directional filters is estimated. Indeed, the maximum value at each pixel (x, y) is picked up as 𝐽(𝑥, 𝑦) = 𝑀𝑎𝑥𝑖 ∈{1,2,…,𝑁+1} (𝐼𝑖 (𝑥, 𝑦)) (5) If J(x, y) exceeds a threshold 𝑇𝐻 = 0.06 it is classified as hair. Otherwise, the pixel is kept unchanged. Fig. 3. Illustrates the process of converting the original colour image (a) into blue component image (b) and then detected hair are removed (c).
(c)
Figure 2. Pre-processing for artifacts removal; (a): original image, (b):blue component image, (c ): after artifacts removal
b) Hair detection Hair detection and removal is a big challenge for extracting the pigment network from images especially when hair has similar linear shape to the pigment network. This similarity can cause incorrect detections. It is worth noting here that the technique used for detecting hair is the same used for pigment network extraction. That is, the directional Gabor filters are applied followed by finding the local maximum at each pixel location. However, the parameters of the Gaussian filters used for each stage are different [13]. The hair detection algorithm uses a bank of 64 directional filters to perform the detection. Let ℎ𝜃𝑖 (𝑥, 𝑦) = 𝐺1 (𝑥, 𝑦) − 𝐺2 (𝑥, 𝑦)
directional filter. The output of the ith directional filter is given by the following convolution:
2) Pigment network extraction The pigment network has a linear shape and appears similar to hair artifacts; therefore, a similar approach can be used to extract the pigment network image with different parameters and threshold values [see (1) – (5)]. These empirical parameters and thresholds have been taken from [13]. Fig. 4. Illustrates the process of converting the original colour image (a) into blue component image (b) and then extracted pigment network image is displayed in (c). It is worth mentioning that negative values in the extracted pigment network image are set to 0 because they normally represent noise.
(1)
Reflection removal
be the impulse response of a directional filter with angle i where 𝜃𝑖 𝜖 [0, 𝜋], 𝑖 = 0, … 64 and 𝐺𝑘 is a Gaussian filter: 𝐺𝑘 (𝑥, 𝑦) = 𝐶𝑘 𝑒𝑥𝑝 {–
𝑥′
2
2𝜎𝑥2 𝑘
−
𝑦′
2
2𝜎𝑦2 𝑘
} , 𝑘 = 1,2.
(2)
(a)
(b)
(c)
Figure 3. Pre-processing for artifacts and hair removal; (a): original image, (b): blue component image, (c ): after hair detection and removal
The difference between two consecutive filters ℎ𝜃𝑖 and ℎ𝜃𝑖+1 𝜋 is constant and equal to . Therefore, N+1 filters are used. In 𝑁 (2), 𝐶𝑘 is a normalization constant and the values of (𝑥 ′ , 𝑦 ′ ) are dependent on (x,y) by a rotation of amplitude 𝜃𝑖 𝑥 ′ = 𝑥 cos 𝜃𝑖 + 𝑦𝑠𝑖𝑛 𝜃𝑖
3𝑎
𝑦 ′ = 𝑦 cos 𝜃𝑖 + 𝑦𝑠𝑖𝑛 𝜃𝑖
3𝑏
The parameters 𝜎𝑥𝑘 and 𝜎𝑦𝑘 for the first and the second filters are chosen to be the value for first filter is less directional than the value for second filter. These parameters have been empirically obtained and are given by: 𝜎𝑥1 = 20, 𝜎𝑦1 = 6, 𝜎𝑥2 = 20, and 𝜎𝑦2 = 0.5. The mask of the filters used have a dimension of 41 × 41. The difference of Gaussians is used in (1) because it allows a better enhancement of directional structures while removing the effect of the background. The image I is filtered by each
(a)
(b)
(c)
Figure 4. Pigment network extraction; (a): original image, (b): blue component image, (c ): pigment network image magnified by 10
3)
Feature-extraction
Statistical means and standard deviations have been used in this work in order to extract features from the obtained pigment network image. The proposed method suggests that the pigment network image is divided first, into (MM) blocks (M=2, 4), each of them can be used as a sub-image from which the mean and standard deviation are computed. This idea is justified by the fact that abnormal images
7215
(corresponding to skin cancer) normally have affected regions in which the pigment network is irregularly distributed with thickened lines than in normal lesions. Therefore, using statistical values could describe these discrepancies in an efficient way. . The detected hair pixels have been discarded in computing these statistical values. 4)
Classification
The fourth and final stage of the proposed system aims to reach a decision on the skin image type using a machine learning technique. In this work, the extracted features have been used to feed an Artificial Neural Network (ANN) as a binary classifier; i.e. the target will be (0) for abnormal and (1) for normal images respectively. The processes of classification were repeated six times according to the number of selected features and a cross-validation (CV) strategy has been used. A two-layer ANN (a single hidden layer and an output layer) has been used with 120 neurons in the hidden layer and the Radial Basis Function (RBF) as an activation function given by[15]: −(𝑥 − 𝑐)2 ) (6) 𝑟2 Where c is the centre of bell-shaped Gaussian and r is the width.
shows that the combination of standard deviations with statistical means bring improvements for a (22) sub-image partitioning. d) Extracted means features (sixteen blocks) Table I illustrates that after dividing the pigment network image into sixteen blocks, the output accuracy results were improved in all cases. The achieved results from applying means features were 92.5% and scores good results for false negative rate and the best result for the false positive rate among other separated features. e)
Extracted standard deviation features (sixteen blocks) Selected standard deviation features score the best overall accuracy [see figure 5], and the best false negative rate among other selected features (see table I). Extracted means & standard deviation features (sixteen blocks) The selected sixteen means and sixteen standard deviation features were merged together to check how could the output improved. Table II illustrates that these selected group of features scores 93% overall accuracy (see figure 5) and low false positive and false negative rates (see table II).
ℎ(𝑥) = 𝑒𝑥𝑝 (
III. EXPERIMENTAL RESULTS Intensive experiments have been conducted on a dataset of 200 dermoscopic images obtained from [1]. The images are considered to be consisting of two classes’ malignant melanoma and non-melanoma. The classification accuracy has been used as an evaluation measure. It represents the rate of correctly classified images. Also, the False Positive Rate FPR (rate of falsely detecting a normal image as cancer image) and the False Negative Rate FNR (falsely missing a cancer image) have been adopted. As mentioned earlier, different sizes of the feature vector have been used depending on the type of features (mean and/or standard deviation) and the number of sub-images. This is summarized below. a) Extracted mean (four blocks) Four extracted features applied to train the system and the output results were recorded after testing the system. The FNR was very low. On the other hand, the FPR is a much higher although the accuracy was 86% [see table I].
f)
TABLE I.
Number of features FNR FPR Accuracy result % TABLE II.
Number of features FNR FPR Accuracy result %
Selected separated
four and sixteen features
Classification output 4 4 16 Features Features Features (Means) (Stds) (Means) 0.067 0.1 0.058 0.25 0.083 0.062 86.0% 91.0% 92.5%
16 Features (Stds) 0.05 0.075 94%
Selected merged eight and thirty two features Classification output 8 Features 32 Features (Means & Stds) (Means & Stds) 0.075 0.058 0.1 0.087 91.5% 93.0%
b) Extracted standard deviation (four blocks) Standard deviation features show better results than mean features. Both false positive rate and false negative rate were good and the overall accuracy was 91% [see table I]. c) Extracted means & Stds. features (four blocks) Merging means and standard deviation features and using them as eight input features gave 91.5% overall accuracy result and better scores in the false positive and negative rates than the separated features [see table II]. This clearly
7216
Figure 5. Selected subset features and accuracy achieved
[4] In table III, we report the variation of the classification accuracy with 16 standard deviation features by repeating cross-validation tests 20 times where μ and represent the mean and standard deviation of the classification accuracy. It can be seen that the change in performance is insignificant against the number of folds and tests. TABLE III.
μ
[7]
16 Features (Standards deviation) 5 Folds Cross10 Folds Crossvalidation validation 92.9% 93.7% 0.50262 0.59382 IV. CONCLUSION
[8]
ACKNOWLEDGMENT
REFERENCES
[3]
[9]
[10]
[11]
The authors would like to thank Catarina Barata for sharing the database provided by Teresa Mendonc, Pedro M. Ferreira, Jorge S. Marques, Andre R. S. Marcal, and Jorge Rozeira which is acquired at Pedro Hispano Hospital. The provided database has been very useful for this work.
[2]
[6]
Variation of the classification accuracy with repeated cross-validation tests
This paper presents a pigment network-based skin cancer detection system. The proposed system uses existing tools for pigment network extraction and extracts statistical features from the pigment network image. The features have then been used to train a neural network classifier. Experiments carried out on a set of annotated dermoscopy images show that the proposed system offers high classification accuracy with very low false detection rates. Future work should focus on pigment network analysis in order to extract new efficient features to further enhance the results. Testing the system with a different larger dataset is also a future goal.
[1]
[5]
T. Mendonca, P. M. Ferreira, J. S. Marques, A. R. S. Marcal, and J. Rozeira, "PH 2-A dermoscopic image database for research and benchmarking," in Engineering in Medicine and Biology Society (EMBC), 35th Annual International Conference of the IEEE, pp. 5437-5440, 2013. K. Hoffmann, T. Gambichler, A. Rick, M. Kreutz, M. Anschuetz, T. Grünendick, A. Orlikov, S. Gehlen, R. Perotti, and L. Andreassi, “Diagnostic and neural analysis of skin cancer (DANAOS). A multicentre study for collection and computer‐aided analysis of data from pigmented skin lesions using digital dermoscopy,” British Journal of Dermatology, vol. 149, no. 4, pp. 801-809, 2003. N. Smaoui, and S. Bessassi, “A developed system for melanoma diagnosis,” International Journal of Computer Vision and Signal Processing, vol. 3, no. 1, pp. 10-17, 2013.
[12]
[13]
[14]
[15]
7217
A. Chiem, A. Al-Jumaily, and R. N. Khushaba, "A Novel Hybrid System for Skin Lesion Detection." pp. 567-572, 2007. L. Ho Tak, and A. Al-Jumaily, "Automatically Early Detection of Skin Cancer: Study Based on Nueral Netwok Classification." pp. 375-380, 2009. (2014, November). Cancer Research UK. Available: http://publications.cancerresearchuk.org/downloads /product/CS_KF_SKIN.pdf M. Emre Celebi, H. A. Kingravi, H. Iyatomi, Y. Alp Aslandogan, W. V. Stoecker, R. H. Moss, J. M. Malters, J. M. Grichnik, A. A. Marghoob, and H. S. Rabinovitz, “Border detection in dermoscopy images using statistical region merging,” Skin Research and Technology, vol. 14, no. 3, pp. 347353, 2008. J. Glaister, A. Wong, and D. A. Clausi, "Segmentation of Skin Lesions From Digital Images Using Joint Statistical Texture Distinctiveness," Biomedical Engineering, IEEE Transactions on, vol. 61, no. 4, pp. 1220-1230, 2014. H. Ganster, A. Pinz, R. Rohrer, E. Wildling, M. Binder, and H. Kittler, “Automated melanoma recognition," Medical Imaging, IEEE Transactions on, vol. 20, no. 3, pp. 233-239, 2001. K. Norton, H. Iyatomi, M. E. Celebi, G. Schaefer, M. Tanaka, and K. Ogawa, " Development of a novel border detection method for melanocytic and non-melanocytic dermoscopy images." pp. 54035406, 2010. H. Rui, C. M. Queen, and G. Zouridakis, "Detection of Buruli ulcer disease: Preliminary results with dermoscopic images on smart handheld devices." pp. 168-171, 2013. C. Doukas, P. Stagkopoulos, C. T. Kiranoudis, and I. Maglogiannis, "Automated skin lesion assessment using mobile technologies and cloud platforms." pp. 2444-2447, 2012. C. Barata, J. S. Marques, and J. Rozeira, “A System for the Detection of Pigment Network in Dermoscopy Images Using Directional Filters,” Biomedical Engineering, IEEE Transactions on, vol. 59, no. 10, pp. 2744-2754, 2012. M. Silveira, J. C. Nascimento, J. S. Marques, A. R. S. Marcal, T. Mendonca, S. Yamauchi, J. Maeda, and J. Rozeira, “Comparison of Segmentation Methods for Melanoma Diagnosis in Dermoscopy Images,” Selected Topics in Signal Processing, IEEE Journal of, vol. 3, no. 1, pp. 35-45, 2009. M. D. Buhmann, Radial basis functions: theory and implementations: Cambridge university press, 2003.
