Feature and model selection with discriminatory visualization for diagnostic classification of brain tumors

Abstract

Machine Learning (ML) and related methods have of late made significant contributions to solving multidisciplinary problems in the field of oncology diagnosis. Human brain tumor diagnosis, in particular, often relies on the use of non-invasive techniques such as Magnetic Resonance Imaging (MRI) and Spectroscopy (MRS). In this paper, MRS data of human brain tumors are analyzed in detail.

The high dimensionality of the MR spectra makes difficult both their classification and the interpretation of the obtained results, thus limiting their usability in practical medical settings. The use of dimensionality reduction techniques is therefore advisable. In this work, we apply feature selection methods and several off-the-shelf classifiers on various ¹H-MRS modalities: long and short echo times and an ad hoc combination of both. The introduction of bootstrap resampling techniques permits the obtention of mean performance estimates and their variability. Our experimental findings indicate that the feature selection process enhances the classification performance compared to using the full set of features. We also show that the use of combined information from the different echo times is a better strategy for small numbers of spectral frequencies; however, the use of ever greater numbers of short echo time frequencies permits the obtention of many models with similar performance. The final induced models offer very attractive solutions both in terms of prediction accuracy and number of involved spectral frequencies, which are also amenable to metabolic interpretation. A linear dimensionality-reduction technique that preserves class discrimination capabilities is used for visualizing the data corresponding to the selected frequencies.

Introduction

Over the last decade, ML has made significant inroads in the fields of bioinformatics and biomedicine. One particular application area that has attracted the attention of both medical practitioners and data analysts is that of human oncology [1]. In this work we are specifically interested in quantitative information in the form of patients’ biological signals. We analyze data corresponding to different types of human brain tumors, obtained by single-voxel proton magnetic resonance spectroscopy (¹H-MRS), with the purpose of developing reliable tools for the support of medical expert diagnostic decision making. Decisions in this area are extremely sensitive and are usually based on information obtained by non-invasive measurement techniques.

The analyzed data belong to a multi-center international database that contains cases of a number of brain tumor pathologies [2]. MRS provides a detailed metabolic fingerprint of the tumor-affected tissue that varies according to the echo time of the acquisition and can be used to characterize these pathologies. The echo time is a relevant parameter of ¹H-MRS measurement, given that, at short-echo times (SET), some of the metabolites are better resolved—although numerous overlapping resonances exist, making the spectra difficult to interpret [3]. The use of a long echo time (LET) yields less clearly resolved metabolites but also less baseline distortion, resulting in a more readable spectrum.

The available data have been acquired both at SET and LET and bundled into three groups or super-classes as described below; they are scarce and of high dimensionality, making their discrimination a non-trivial undertaking. Therefore, the need arises for the use of dimensionality reduction methods (feature selection and/or extraction) in order to reduce the overall complexity of the problem. We use an entropic filtering algorithm for feature selection as a fast method to generate relevant subsets of spectral frequencies. An in-depth feature selection study is performed, not only in LET and SET data, but in a combination of both echo times. Bootstrap resampling techniques are used to yield mean performance estimates and their variability, and thus a more reliable measure of predictive ability. The combination of feature selection and classification aims at obtaining simple models (in terms of low numbers of features) capable of good generalization.

We report experimental results that support the practical advantage of combining robust feature selection and classification in this application, as accurate classification is obtained with parsimonious and interpretable subsets of spectral frequencies. We also aim to progress in the comparison of performances for MRS data acquired at different echo times, as well as in the comparison of these with data that combine both echo times.

Of special importance in a practical medical setting is the interpretability of the solutions in terms of these spectral frequencies, something that limits the applicability of methods such as PCA or ICA (whose solutions involve weighted combinations of frequencies, instead of individual frequencies). Moreover, even if interpretable by mere inspection of the involved features, the final selection of spectral frequencies may still provide few clues about the structure of the classes (tumor types). In this medical context, data visualization in a low-dimensional representation space may become extremely important, as it would help radiologists to gain insights into this complex and highly sensitive domain. A linear dimensionality reduction technique that provides a data projection—while preserving the class discrimination achieved by a classifier—is also used in our study. The goal of combining feature selection and visualization is to increase the intuitive interpretability of the classifier results.