Multi objective evolutionary algorithms and their application to financial portfolio optimization (PhD)

Abstract

The invention of the microscope has revolutionized scientific research and industry,enabling the unraveling of the microscopic world. Microscopy images play an essential role in various fields of biomedicine, engineering, and physical sciences. For medicine, examination of microscopy images provides objective support for improving assessment of various disorders such as breast cancer, prostrate cancer, lung cancer, autoimmune disorders including addison’s disease, celiac disease etc. However, in clinical applications, it is usually time consuming to examine an image manually. Moreover, there is always a subjective element related to the pathological examination of an image. This can result in a potential risk of achieving lower sensitivity/specificity in the manual examination. Computer-aided clinical image analysis has attracted interest from both computer science and medical researchers due to its potential to surmount the challenges associated with the subjective examination of microscopic images. Computer-assisted diagnosis (CADx) systems aim to assist medical physicians for making diagnostic decisions with computers, so that instead of spending time for cases where decisions can be made in a straight forward manner, the pathologist can focus more on the ambiguous cases. The goal of this thesis is to develop new algorithms for classification of microscopy images by employing ensemble and deep learning techniques. The work explores different classifier ensembles, a variety of feature representations, and methods for feature selection. These are considered in context of both traditional and deep learning paradigms. The aspects are considered for two different microsopic image analysis problems. In the first part of this thesis, the aim has been to develop novel methods to automatically classify positive staining patterns of HEp-2 cells which are widely used in diagnosis of autoimmune diseases. Firstly, the idea of extracting a relatively small set of class-specific features which in turn will be motivated by the visual morphological characteristics of each class has been proposed. The primary objective of this feature extraction approach is to utilize morphological properties of stained-cell regions i.e. those features which capture the patterns that are apparent over and above visual traits. When one considers the fact that subsets of ‘good’ features from the original dataset are found to be sufficient in providing discriminative information regarding the data classes, the second part of this study has focused on identifying irrelevant and redundant features (attributes). This has been achieved by an idea that a classification model built only with a specific subset would have better predictive accuracy than a model built with a complete set of features. Besides the considerations which come with existing feature selection (FS) methods for this task, there search work has proposed a hybridization technique which combines filter-based FS techniques with automatic select feature subset (robustness of feature ranking) thereby accessing classification performance pre/post such feature selection. The third part of this study has explored the effectiveness of CNN features along with the proposed feature set in a heterogeneously designed committee of network sand classifiers where diverse members have been selected based on an information theoretic measure. Experimental studies have confirmed the proposed work to provide encouraging results and highlight the role of class-specific features across various classification frameworks. The second problem of microscopy image analysis addressed in the thesis relates to imagery of histopathology. This part of the thesis investigates images of breast cancer histopathology based on computer-assisted image analysis. In the first study of this part, different frameworks for integrating multi-layered deep features extracted from fine-tuned deep networks are investigated.The frameworks designed take into consideration the layers sequential and independent nature. It is demonstrated through various experiments that the proposed multilayer feature fusion actually outperforms the baseline network, as well as classification using only the highest level features. It is also shown that the proposed approach outperforms most state-of-the-art methods of classification. The notion of layer selection is explored in the second study, considering that not all layers can contribute to decision making, as many of them learn similar representation (small variance). In addition, since histopathological images show a high degree of variability, useful information is often obtained at different levels of optical magnification to make the correct diagnosis. A multi-scale model is proposed in this regard to combine scores from images at different magnification. From the experimental studies, it is concluded that if one includes scores integrated across different magnifications, more beneficial decisions can be made.