Pattern Classification of Melanoma by Local Features Using BoF Based Spatial Encoding

-Melanoma is the direst form of skin cancer. Early detection of cancer is a very critical issue in today’s dematologic practice. Feature extraction allows representing the content of the image as perfectly as possible. In this paper, Bag of Features based Supervised Spatial Encoding of Feature extraction is proposed. Low Level Images and their Latent Features induce codebook of features. Scale invariant Speeded up Robust Features (SURF) technique is used for feature point detection in Low Level Image representation. Since the system considers the entire image as lesion, it also recognizes clustered or patched lesion. It uses l*a*b color space for describing color intensities. The patterns so detected are classified using multi-SVM classifier. The proposed cluster based system provides the classification accuracy of 95.075% and sensitivity, specificity rates as 94.07% and 95.5 respectively.


II RELATED WORKS
There are several systems for the identification of melanoma in dermoscopy images. Most of the papers focus to identify the lesion and its malignancy. Papers that focus pattern classification use global methods to classify the pattern. Global methods consider color, shape and texture features [8]- [12]. The standard approach in automatic dermoscopic image analysis has usually three stages: 1) image segmentation; 2) feature extraction and feature selection; and 3) lesion classification. Yogendra et.al., [14] presented A new approach for skin cancer classification by which it detect features of a digital image by decomposing the images into different frequency sub bands using wavelet transform. The classification methodology is based on probabilistic neural network and clustering classifier and also classify that whether the given input image is cancerous or non cancerous. Barata et.al . [15] presented a bag of features approach for the classification of melanomas in dermoscopy images a comparison between color and texture feature. The feature extraction is represented by a two dimensional matrix known as GLCM. The color features are computed by color histograms. The classification is based on K-nearest neighbor and Support Vector Machine (SVM) and the result shows that color feature outperforms the texture feature. Tanaka et al. [17] presented Pattern classification of nevus with Texture Analysis by considering Co-occurrence matrix on images. Melanoma here is diagnosed by features of specific shape, color and texture. It classifies three patterns by texture analysis: homogeneous pattern, globular pattern and reticular patterns. The tumors are not properly extracted in some images because the area of the tumor was very small and also the contour of the tumor was blurred. Lucia et.al. [16] presented a novel hierarchical classification system. The segmentation is done by using levelset framework. The texture features are extracted from Generalized Co-occurrence Matrix (GCM). The classification methodology is based on K-NN classifier and it classifies the dermoscopy image into melanoma. Aurora saez et.al. [18] presented Model based methods of classification of global patterns in dermoscopic images. The segmentation is done by using level-set algorithm. The texture features are extracted by Markov Random Field (MRF).The classification methodology is based on Gaussian model, Gaussian mixture model and bag-of-features model to classify a pigmented lesion into three categories: globular, homogeneous and reticular. The BoF approach use MRF features of local patches sampled from the training set and the overlapping patches are extracted from an image. This paper uses clustering based melanoma pattern recognition system using spatial encoding with BoF approach for feature extraction. It considers the whole image into Low Level Image (LLI) representation . They are done by BOF feature selection by SURF for that is denoted as . Then convert into vector space by encoding. Region based Statistical Region Merging (SRM) technique [22] [23] is used for segmenting the lesion from the skin .It uses Multi-SVM for pattern classification. The organization of this paper is as follows: Section II describes aim and objectives of the paper. In Section III, the proposed system overview is discussed. Section IV discusses the methodology adopted. The experimental results are evaluated in Section V and Section VI concludes the result and expresses its future scope.

II RESEARCH AIM AND OBJECTIVES
The purpose of the study is to design and develop a system on melanoma pattern recognition area that focuses on early detection of the malignancy level and patterns associated with skin cancers cells and also have less dependency on medical experts or dermatologists. The objective of the paper includes: i) To develop an algorithm for the pattern classification of melanoma by local features using BoF based spatial encoding. ii) Scale and rotation invariant algorithm also solves the recognition of clustered lesion problem iii) To ease and early treatment of skin cancer, the system proposes the melanoma pattern recognition system to identify the distribution of patterns at an earlier stage and its pattern type.
III SYSTEM OVERVIEW The analysis and classification of cluster patterns on melanocytic skin lesion is an object recognition system with specific lesion properties. Given an input image, the system first extracts Low Level Image (LLI) using SURF [24] for the whole image lesion i.e) represent the whole image into BoF representation that is denoted as . Then convert into vector space by encoding. Convert whole image into N N low level image and n n window size. Embedding BoF scattered spatial features into embedded feature set. Classifier used here is Multi-SVM. The following figure depicts the overall design of the BoF based Spatial encoding. The logical steps of the algorithm as follows:

Input : Lesion Output: Melanoma pattern
Step 1 : Read Training database Step 2: covert image from rgb color space to lab color space.
Step 3: Apply PCA filter for noise removal Step 4: Apply SURF feature point detection to the whole image for LLI extraction and coding.
Step 5: Pooling and Embedding feature Vectors Step 6: Read the test image and goto step 1 Step 7: Apply Multi-class SVM for classifying the patterns of the melanoma

IV METHODOLOGY
Differentiating the melanoma patterns from the skin lesion needs effective preprocessing, segmentation, feature extraction and classification techniques. This system mainly focuses on Texture analysis phase.

A. Pre-processing
Image pre-processing is an essential step for dermoscopic images which eliminates noises, leads to increasing accuracy level of the pattern recognition system. Thus, it requires some preprocessing techniques for image enhancement and restoration. Since color information plays a significant role in the analysis of dermoscopic image, the input image here in RGB color space is directly transformed into a LAB color space. And also has the following advantages: (i) decoupling luminance and chromaticity information, and (ii) Achieving invariance to different imaging conditions such as viewing direction, illumination intensity [19]. The advantage of using Lab color space is that it yields uniform color space since Lab is perceptually linear and also the luminance factor L for all pixels is constant. This will reduce the data size and computation time. The second step is to remove the artifacts from the image. This has been achieved by applying PCA filter on the image database followed by image expansion using contour dilation technique. PCA is used for dimensionality reduction of color and texture space [20].

B. Melanoma Pattern observation by Local features using BoF based Spatial encoding
As the main aim of the paper is to classify the clustered lesion into patterns, it focuses the entire region. The reason behind this decision is that some lesion images come in clusters. The paper represents images into Low level images. They are extracted by Bag of Features (BoF) representation using SURF techniques. Single scale spatial encoding of the above approach provides better recognition results compared with other clustering technique. Supervised Spatial Encoding of feature analysis framework is shown in the

Step 1: Low Level image (LLI) Extraction
The first step is the low level image extraction. Let L be the total number of training images L=305 for our system. Y denotes the class labels Y=1,2……C.Y=4 for our system. X is the collection of labeled image denoted as i=1,2….L. Given an image x, partition the whole image into NXN sub regions and extract low level image (LLI) descriptor using SURF for the whole image lesion. i.e) represents the whole image into BoF representation that is denoted as . This system uses single scale.
Step 2: Coding The second phase is the feature coding. A coding step encodes each LLI via non-linear feature mapping into vector space. This phase is induced by codebook of image features.
Step 3: Pooling The third step is the feature pooling, It concatenates the feature codes into a single feature vector. Let the lesion X into N N low level image regions and the window Size be n regions in M-dimensional latent space. Here N=5 and n=2. For sliding window consisting of k= regions { , … … represents BoF of sliding window regions, vector denotes concatenation of its feature vector (FV) as: Where | | is the BoF representation of region w.

Step 4: Window Embedding
The fourth step is the window embedding. The M-Dimensional Embedding for window the window is = where G . | | and taking concatenation of all gives Step 5: Image Embedding The fifth step is the image embedding. It concatenates latent embedding of all sliding windows into a single vector defined as: Where h(.) = tanh . . Then the final image-level latent representation is the result of the second projection ≡ (4) where F∈ . .

C. Classification
Even if the analysts were using number of classification techniques for melanoma prediction like ANN classifier, Ada-boost classifier, K-NN classifier and SVM classifier, the experiments show better results for SVM classifier. Since the pattern includes Homogeneous, Globular, Reticular and Multi-Component, the paper uses multi-class SVM for classification.

V EXPERIMENTAL RESULTS AND EVALUATION A. Experimental data
The experiment is carried out with 305 dermoscopic images taken from Interactive dermatology atlas, opticomdataresearch of four classes (Globular, Homogenous, Reticular and Multi-component patterns ) is n=4.So each image is divided into 9 sub regions and 4 possible window sliding regions. Initially, SURF based representation considers the Gaussian pyramid octave is 8 that selects strongest features. Upright or rotated feature flag is set to recognize rotated images. The system chooses randomly chosen SURF reticular, multicomponent).The images are in JPEG formats and the images are resized into 256x256 pixels. The system is validated by leave one out approach.-For evaluating the results, MATLAB R2012a software was used and sample data inputs as follows in Fig 2. The perf terms are Since the feature extraction operation depends on various parameters, the algorithm explicitly vary the parameters in order to achieve better results. Initially the whole image is partitioned into 2x2 sub image and finds SURF feature points for each sub region that forms 4 bags of cluster points. The number of SURF feature points varies as 50 weak points, 100 strong points and 200 strong points. Coding in the algorithm that form a codebook for each clustered region and BoF of sliding window concatenates the featured codebook. The algorithm tested up to 5x5 dimension of sub image size with varying feature points. Table I summarizes the results obtained on testing of dermoscopic image database with varying sub image levels and SURF feature points. From the observation, the average melanoma patterns are significantly classified when computing on the set of training images with increasing sub image levels. In particular, reticular and Multi-component patterns tend to produce better classification results due to the presence of pigmented network and associated color combinations. The clustered code book contains more information on feature point variations and its discriminations. In the feature extraction and analysis step, the single vector thus formed with window embedding have represents the distribution of small scale features within the interest point neighborhood. This can also be a reason to have better accuracy for reticular and Multi-component patterns and produce significant results for other patterns too. It is also noted that the algorithm perform almost 98% accuracy on working with clustered lesions. Finally, sub image was fixed to L3 (5x5) and set the feature point as 200 in order to achieve the trade-off between complexity and size. Table II provides the results of all performance metrics that considered. The results thus found shows that the proposed method gives better discriminations among various patterns when compared with other clustering techniques. Table III shows the results of existing clustering techniques and the proposed one. In paper [9] an image is modeled by an MRF on the l*a*b color space and are divided into 81x81 patch size. The cluster code book contains MRF features with n-dimensional vector per class produces the accuracy as 72.91% and sensitivity, specificity rates as 80.63% and 74.836% respectively. Clustering based feature extraction used in [15] follows color and texture based feature points as melanoma discriminations. In this BoF based technique, key points are of regular grid size ∆x∆ in the image domain. By varying the bins of histogram as 15,16…50, the algorithm produces the accuracy as 91% and sensitivity, specificity rates as 93% and 88% respectively . From all the above discussions, the proposed clustering technique produces substantial improvement on lesion images over other clustering techniques.

V CONCLUSION AND FUTURE SCOPE
The proposed early detection of the lesion pattern recognition system using BoF based Supervised Spatial Encoding of clustered technique provides better results when compared with other existing clustered techniques. Most of the authors focus on the lesion malignancy. But, this pattern recognition system is used to find the melanoma types. Since the Spatial Encoder considers the whole image as lesion, it also recognizes clustered or patched images. Low Level Images using BoF representation provides vector features. Vector features so formed are concatenated and embedded in the codebook of features. Finally, the features on the codebook are classified using multi-SVM classifier. The performance of the existing methods shows that they provide better results for less pattern types and for less data elements. The system works on clustered images also. Scale and rotation variant images are also tested. This system provides 95.075% accuracy, 94.07% sensitivity and 95.5% specificity rates on an average for four classes and the database uses 305 data elements. This system uses the color and texture features in spatial domain. In future, the system aims to enhance the qualitative rates by considering geometric features. This can achieve by normalizing the feature sets.