Abstract
An automated method to detect and analyze the granular parakeratosis is presented to improve diagnosis which will lead to the exact treatment. Granular parakeratosis (also known as "Axillary granular parakeratosis", "Intertriginous granular parakeratosis", and more recently, "Zombie Patch") is a cutaneous condition characterized by brownish-red keratotic papules that can coalesce into plaques. Granular parakeratosis (GP) is a rare disorder of keratinization with a distinctive histology wherein parakeratosis with retention of keratohyaline granules is noted in the epidermis. Granular parakeratosis often presents as pruritic hyperpigmented or erythematous patches and plaques in intertriginous areas, more commonly in women than men. Here, we present a novel approach for automatic segmentation and classification of skin lesions. This is focused towards the development of improved Automatic Lesion Detection System (ALDS) framework for segmenting skin lesions and later Support Virtual Machine (SVM) is applied for the classification of the segment.
Index Terms — Segmentation, Automatic Lesion Detection System, Skin lesion, Granular parakeratosis, Support Virtual Machine
INTRODUCTION
Granular parakeratosis is a rare, idiopathic and benign skin disease that presents classically as erythematous to brown hyperkeratotic papules that can coalesce into plaques. The term "granular parakeratosis" is now used to describe not only the skin condition, but also a distinctive histological reactive pattern on biopsy specimens that are either regarded as the disease itself, or as an incidental finding.
In the last decades, the digital images produced by educational, medical, industrial, scientific and other applications are used to diagnose the various diseases. The drastic growth of digital electronics industries has posed many challenges in dealing with huge amount of image data. The management of the expanding visual information has become a challenging task.
Image Segmentation is one of the important issues occurred in times of before computer visualization. The fundamental objective of image segmentation is to segment a picture into its constituent areas and accordingly, the action of handling the image can be essentially diminished. These separations are the image objects that claim related texture [1]. The image segmentation brings about an arrangement of areas that mutually cover the image totally or deliver an arrangement of contours separated from the image. The pixels of image areas are connected through its trait or figured properties like color, intensity and texture. The same characteristics make neighboring regions to be different. The most recurrent issue related with image segmentation is the requirement of an integrity measure that can independently assess its functioning. The reason for this trouble is the absence of total ground truth because of various manual segmentations of a similar image. It is the issue of apportioning an image into its constituent parts. In carefully picking a segment that highlights the part and main properties of every segment, we get a solid description of an image regarding its valuable parts. Contingent upon the end application, the issue of segmentation can be subjective or objective
An effective image processing system can be potentially used to segment the granular parakeratosis images. The authors of [2] created a model which can characterize the pigmented skin lesion. Users can query the database by feature attribute values like shape and texture, or by synthesized image colors. Gnanasigamony [3] developed a system for retrieving skin lesion images based on shape similarity. Alfonso Baldi [4] presented a CBIR (content-based image retrieval system for dermatoscopic images. Their approach includes image processing, segmentation, feature extraction (colour and textures) and similarity matching. Classification methods range from discriminant analysis to neural networks and support vector machines. So based on the desired segmentation and classification of the experimenting dataset, the system may opt the Support Vector Machine.
Problem statement
The most fundamental and challenging tasks in digital image analysis is segmentation, which is the process of assigning pixel-wise labels to regions in an image that share some high-level semantics, hence the term “semantic segmentation”. In skin lesion segmentation, the goal is to assign pixel-wise labels to regions in dermoscopy images that represent skin lesions, such as granular parakeratosis, melanoma, seborrhoeic keratosis or benign nevus.
The segmentation of Skin lesion is challenging due to a variety of factors, such as variations in skin tone, uneven illumination, partial obstruction due to the presence of hair, low contrast between lesion and surrounding skin, and the presence of freckles or gauze in the image frame, which may be mistaken for lesions. A successful lesion segmentation technique should be robust enough to accommodate these variabilities.
Skin lesion segmentation is a widely researched topic in medical image analysis. Until recently, most skin lesion segmentation approaches were based on hand crafted algorithms [5–6]. These approaches require carefully designed pre-processing and post-processing approaches like hair removal, edge-preserving smoothing and morphological operations. Hence, there is a great need for robust methods that process with the interpretation of huge amounts of data with greater accuracy. To overcome these difficulties in clinical diagnosis, segmentation of medical images provides the potentiality for increasing the diagnostic accuracy.
Literature survey
In this section, we review some of the primary techniques available in literature for medical image segmentation. In the recent years, various schemes for processing medical images appeared in literature. Researchers have developed many schemes and techniques for segmenting and characterizing the medical images. The use of segmentation is to partition an image into strong correlated parts with “area of interest” in the image. Image segmentation ends, when the object of focus is separated. Segmentation can be classified as complete and partial. Complete segmentation consequences in a set of disjoint regions corresponding absolutely with input image objects, whereas in partial segmentation, resultant regions do not match directly with input image. Usually, image segmentation is treated as a pattern recognition problem as segmentation requires classification of pixels. In medical imaging automated description of different image components are used for analyzing anatomical structures such as skin lesions, bones, muscles blood vessels, tissue types, pathological regions(like cancer, multiple sclerosis lesions) and for dividing an entire image into sub regions.
In [7], the authors segmented lesion areas using region growing method. Later, Color and texture features are extracted to represent segmented lesion areas. Then the classification is performed with SVM, KNN and fusion of SVM and KNN Classifiers. It is observed that the performance of the system has decreased due to the fact that feature selection methods are not used to select good features
Garrett Nelson MD, Mary H. Lien MD, Jane L. Messina MD, Sonali Ranjit BS, and Neil Alan Fenske [8] examined Paraffin-embedded, H&E-stained sections of a punch biopsy and demonstrated slight acanthosis, and a sparse superficial perivascular dermal lymphocytic infiltrate. The most striking feature was marked – thickening of the stratum corneum, which was hypereosinophilic.
Ding CY, Liu H, Khachemoune A [9] proposed the reappraisal of GP as a reactive pattern, rather than a distinct entity
The authors of [10] carried out segmentation process to extract the lesion area from the selected image. The features are extracted from the segmented image based on the Gray Level Co-occurence Matrix (GLCM) method. GLCM is a method used to extract texture features from a gray level image. The various skin lesions are classified based on their texture features using Support Vector Machine (SVM) and K Nearest Neighbor (KNN) classifiers. This system extracts only texture features of the selected image to perform classification. Other features such as color, shape etc can be extracted to get better classification
An automated skin cancer diagnostic system is proposed based on self-advising SVM in [11]. Self-advising SVM uses information generated from misclassified data in the training phase and thus, improves the SVM performance by transferring more information from the training phase to the test phase. It is also taken into account that diverse range of features can be extracted from skin images using state of the art feature extraction methods to enhance the classification performance of the classifier. But the system fails to develop more reliable diagnostic system.
In [12], SVM has been implemented for classification of benign from malignant skin tumor. MATLAB package is used to implement the software in the current work, these feature were carried out to generate training and testing of the proposed SVM. But this paper concludes that there are some possible factors to improve the accuracy of detecting malignant melanoma
A.A.L.C Amarathunga [13] presented a development of a skin diseases diagnosis system which allows user to identify diseases of the human skin and to provide advises or medical treatments in a very short time period. They used various classifiers to calculate and evaluate the accuracy level of our system. Multi-Layer Perceptron (MLP) and J48 are main classifiers we used. This application is applied only for three skin diseases. They are Eczema, Impetigo and Melanoma. They developed this only for windows application. It is not yet develop for smart phones like Android, IOS and etc. And another thing is when capturing the image the distance between camera lens and affected skin is 5cm.