Abstract
The high number of exams that is done in healthcare institutions increases the medical doctors’ workload, leading to poor working conditions and the increase of wrong diagnoses. As consequence, an automatic system that can help medical doctors in diagnostic tasks is of major interest to any healthcare institution. The chapter proposes an Image Based Classification Platform suitable to help Medical Doctors diagnosing breast cancer, based on mammograms, i.e., to detect if a tumor is present in the image. The platform is twofold, i.e., in the first part the image descriptors are extracted from the image using image-processing algorithms. The obtained descriptors are used in the second part. The second part is related to classification, where computational intelligence methods are used to classify a given image, based on the descriptors obtained in the first phase. Texture analysis based on co-occurrence matrices are applied to obtain the descriptors from the MIAS database of mammograms. From these descriptors, fuzzy models, neural networks, and support vector machines are successfully used to classify the mammograms and obtain a diagnosis.
Top1. Introduction
This chapter presents a software tool to help medical doctors in diagnostic tasks, using medical images. This software tool is divided in two complementary parts, which can work together or separately; the first is dedicated to image processing and the second to classification tasks. In this chapter, both parts work together to perform classification tasks based on information obtained from images.
The first part of the system gives the image descriptors of the captured image. The algorithms implemented in the software tool are the wavelet decomposition for texture and color (Alexandre & Pinto, 2006), gray-level co-occurrence matrix (Haralick, 1973), quadtree decomposition (Ramalho, Caldas Pinto and Marcolino, 2000) and Markov random fields (Yu & Cheng, 2003). These algorithms give a large number of features that are used in the second part of the system. The descriptors to be used can be chosen by the user. The second part of the system uses the image descriptors obtained in the image processing part of the system. The computational intelligence systems implemented in the platform are based in Fuzzy Models (Takagi & Sugeno, 1985; Zadeh, 1973; Mamdani, 1977), Neural Networks (Zhang, 2000) and Support Vector Machines (Vapnick, 1998). The software tool described in this chapter is applied to Breast Cancer diagnosis, based in mammograms. The tool classifies the mammograms in two classes: with tumor and without tumor. From the diagnostic data, mammograms, the software tool can propose to the medical doctor a possible assessment.
Breast Cancer is a major cause of death amongst women in the world (Ferlay, 2010). In the world, 22.9% of Cancer diagnoses are Breast Cancer, which causes 13.7% of Cancer deaths amongst women (Ferlay, 2010). In the European Union (EU) this reality is been tackled to increase the survival rates of women with Breast Cancer diagnose, by governments. In the EU, the overall spending of Breast Cancer care amounts to about 0.5-0.6% of total health care expenditure (Organisation for Economic Co-operation and Development (OECD), 2010). This effort has been rewarded with the decrease or stabilization of the death rates in the EU, despite the increase of incidence rates of Breast Cancer over the past decade (OECD, 2010). These facts show the increase of the survival rates due to earlier diagnosis and/or better treatments (OECD, 2010).
According to the OECD Health Data (OECD, 2010) the survival rate is 79.4%, ranging from 61.6% in Poland to 88.3% in Iceland, with only 51.4% of the women (with 50 to 69 years) screened in the OECD countries, ranging from 10.7% in Turkey to 84.4% in Finland. This last fact, i.e., the percentage of screened women, is of great importance to detect Breast Cancer in the early stages, which increase can contribute to diminish the mortality rates. The governments screening programs based on mammograms have been discussed by (Cady & Chung, 2005) concluding that its benefits overtake its drawbacks, suggesting that its usage should be expanded.
Key Terms in this Chapter
Computational Intelligence: A set of methods suited to tackle the uncertainty, imprecision, of real systems, mostly nonlinear. These methods approximate real systems to obtain low computational cost, robust and tractable models.
Neural Network: A network of mathematical neurons that simulate the human brain, therefore approximating the behavior of nonlinear systems.
Support Vector Machine: A classifier that separates the data in several groups by hyperplanes, in the feature space. It is closely related to neural networks.
Grey-Level Co-Occurrence Matrix: A matrix that defines the distribution of co-ocurring grey-level values within an image for a given offset vector.
Computational Intelligence: A set of methods suited to tackle the uncertainty, imprecision, of real systems, mostly nonlinear. These methods approximate real systems to obtain low computational cost, robust and tractable models.
Neural Network: A network of mathematical neurons that simulate the human brain, therefore approximating the behavior of nonlinear systems.
Wavelets: Are “little waves” with specific properties that are the base for the decomposition of nonlinear functions.
Markov Random Fields: Are multi-dimensional random processes, mainly Gaussian in images that can represent images.
Support Vector Machine: A classifier that separates the data in several groups by hyperplanes, in the feature space. It is closely related to neural networks.
Takagi-Sugeno Fuzzy Model: A finite set of fuzzy piecewise linear relations to approximate a nonlinear system. This model is usually obtained from measured input-output data from the nonlinear system.
Markov Random Fields: Are multi-dimensional random processes, mainly Gaussian in images that can represent images.
Grey-Level Co-Occurrence Matrix: A matrix that defines the distribution of co-ocurring grey-level values within an image for a given offset vector.
Takagi-Sugeno Fuzzy Model: A finite set of fuzzy piecewise linear relations to approximate a nonlinear system. This model is usually obtained from measured input-output data from the nonlinear system.
Wavelets: Are “little waves” with specific properties that are the base for the decomposition of nonlinear functions.