' Machine Vision Approach in the Classification of Breast Cancer Using Texture Analysis''

Abstract

The main objective of this research is to develop a diagnostic system, based on standard or objective parameters rather than non-standard or subjective parameters, which are already being employed by radiologists, for the classification of abnormalities present in mammograms, as benign or malignant. Radiologists differentiate biological behavior of these abnormalities on the basis of visual parameters such as size, shape and boundaries of the tumors. A benign tumor has small size, well-defined margins and homogenous texture, whereas, a malignant tumor typically has larger size, poorly margined and heterogeneous texture. Due to the limitations of human perception all these parameters become subjective, which cause a high risk of misinterpretation, inter and intra- observer variation, for correct decision. Considering all these factors, development of a compact system is required; (i) to accurately classify malignant and benign abnormalities within a reasonable time and cost, (ii) to increase diagnostic consistency by providing an objective (rather than subjective) evaluation. To meet the objectives a CAD system has been developed and employed on two types of data; (a) test data and (b) experimental data. In test data, five wheat varieties are differentiated by this system. For this purpose five types of 77 statistical textural features, which may be grouped as; first order (histogram) features, second order (GLCM) features, higher order (GLRM) features, autoregressive features, and gradient matrix based features are calculated from ROIs (8x8) (16x16) (32x32), and (64x64) under, μ±3σ and 1-99% normalization conditions by using MaZda software. The most relevant features for each size of ROI are selected by three approaches; Fisher’s Co-efficient, Probability of Error plus Average Correlation Co-efficient, and Mutual Information Co-efficient. In this way the most relevant 10 features are selected by each method. We received very poor results when data analysis capability is verified on the basis of 10 features selected by each method for each size of ROI except (64x64), by three multivariate techniques; PCA, LDA, and NDA under both normalization conditions, by a software „B11‟, integrated with MaZda.To improve the results, a set of 19 features is obtained by merging the features selected by each approach. An excellent clustering result with an accuracy of 99.67% is received, when data of these 19 features extracted from ROI (64x64) under 1-99% normalization, is deployed to NDA projection space. By using supervised classification approach, artificial neural network (ANN) the system is trained and tested on the basis of 70% and 30% of input data respectively. We received an accuracy of 99.90% and 93.11% in training and testing phase respectively. On the basis of results for test data analysis, it is concluded that the proposed CAD system produces the best result for large ROI window size when a combined set of features is deployed in NDA projection space. The experimental data (mammograms) is analyzed under these settings. The mammographic data is consisted of two types of images, Craniocaudel (CC) and Medioletral Oblique (MLO) view images. Unlike to other researchers in this work both types of images are considered in separate sections. As the experimental data (mammograms) has fine and micro-texture, so, initially abnormal regions in CC view images, marked by radiologists, are tried to be analyzed on the basis of combined set of features (discussed above for the case of test data), extracted from ROI (8x8) under above mentioned both normalization conditions. As NDA approach based on ANN classifier and a number of options are available on „B11‟ software to configure this classifier. Data analysis capability of selected features under different architectural settings of ANN on the basis number of neurons in input hidden layer and learning rate „η‟ in NDA projection space is tried. Then the classifier is trained and tested on the basis of architectural settings for which the best clustering is received, by splitting data in 70/30 ratio respectively. For ROI (8x8) we received testing accuracy of 91.18% when the classifier is configured with 2 neurons in hidden layer and learning rate is set at 0.15 when the features are extracted under μ±3σ normalization condition. For same of size of ROI when features are extracted under 1-99% conditions, the best testing accuracy of 88.44% is obtained for same architectural settings (2 neurons and η=0.15). As the performance of the system for ROI (8x8) is not satisfactory, so, we tried to classify same data by extracting features from ROIs (16x16) under the both normalization approaches. Following the same procedural steps we received an accuracy of 92.56% for testing phase, when the classifier is configured with 2 neurons in hidden layer at learning rate 0.20 and the features are extracted under 1-99% normalization. We received excellent classifier testing result with an accuracy of 97.55% when the classifier is configured with 3 neurons in hidden layer at learning rate 0.15 and the features are extracted by applying μ±3σ approach. For MLO view images we obtained a testing accuracy of 84.41%, when the classifier is configured with 5 neurons in hidden layer with learning rate is set at 0.22 and the data is normalized by μ±3σ approach.