A Content-Based Image Retrieval System for Diagnosis and Detection of Skin Cancer Using Self-Organizing Feature Maps

Abstract

A large number of medical images with skin blisters are stored on distributed and centralized servers and are referred for knowledge, teaching, information, and diagnosis. The Content-Based Image Retrieval (CBIR) system is used to locate images in vast databases. Images are indexed and retrieved with a set of features. The CBIR model, on receipt of query, extracts same set of features of query, matches with indexed features index, and retrieves similar images from database. Thus, the system performance mainly depends on the features adopted for indexing. Features selected must require lesser storage, retrieval time, cost of retrieval model, and must support different classifier algorithms. Feature set adopted here consists of Local Binary Pattern (LBP) and the power spectrum coefficients from Discrete Fourier Transform (DFT), which provides support to improve the performance of the skin cancer detection system. The chapter briefs the strength of LBP values in fusion with the DFT coefficients for categorizing blisters as highly cancerous (malignant) and noncancerous (benign) from the database. The medical images of skin cancer are taken for analysis from DermIS and DermQuest. The results presented in this chapter are obtained by using a clustering technique like Self Organizing Maps (SOM), which uses the distance measures like L1 or Manhattan distance measure (L1), Euclidean distance measure (L2), d1 distance measure, and Canberra distance measures, respectively. The results prove to have a good prospectus for fusion of LBP features and variance values apart from considering DFT coefficients for clustering. The Identification Efficacy (IE) is in the range of 85% to 99.5%. Melanoma is a deadly form of skin cancer. This serves as the cause for the development of a highly cancerous tumor on the skin. The dermatological photographs are used to detect the skin cancer. The objective of this project is to develop a structured scheme to analyze and evaluate the probabilities of melanoma with the help of a typical user-friendly camera. The novelty of this project is that it has a well-developed strategy for skin cancer detection that uses high-performance image-based machine learning algorithms to extract the Fourier coefficients followed by the formation of LBP values for the region of interest, the lesions on the skin surface. This method uses images from the open source database like DermIS and DermQuest. The power spectrum extracted using Discrete Fourier Transform of the MR images is evaluated and they play an important role in increasing the sensitivity for identifying the highly cancerous blisters on the skin. These power spectrum coefficients are used as distinct input features that are used as a dataset for training SOM to detect and identify the highly cancerous lesions on the skin surface.