Voxel-Wise Feature Selection Method for CNN Binary Classification of Neuroimaging Data

Domenico Messina,Marco Salvatore,Marco Aiello,Pasquale Borrelli,Paolo Russo

doi:10.3389/fnins.2021.630747

Abstract

Voxel-wise group analysis is presented as a novel feature selection (FS) technique for a deep learning (DL) approach to brain imaging data classification. The method, based on a voxel-wise two-sample t-test and denoted as t-masking, is integrated into the learning procedure as a data-driven FS strategy. t-Masking has been introduced in a convolutional neural network (CNN) for the test bench of binary classification of very-mild Alzheimer’s disease vs. normal control, using a structural magnetic resonance imaging dataset of 180 subjects. To better characterize the t-masking impact on CNN classification performance, six different experimental configurations were designed. Moreover, the performances of the presented FS method were compared to those of similar machine learning (ML) models that relied on different FS approaches. Overall, our results show an enhancement of about 6% in performance when t-masking was applied. Moreover, the reported performance enhancement was higher with respect to similar FS-based ML models. In addition, evaluation of the impact of t-masking on various selection rates has been provided, serving as a useful characterization for future insights. The proposed approach is also highly generalizable to other DL architectures, neuroimaging modalities, and brain pathologies.

Highlights

During the last decade, technological advancements and the availability of large amounts of labeled data (Aiello et al, 2019) fostered neuroimaging research’s development (Traverso et al, 2020)
The results were confirmed for classical crossvalidation metrics, where fs_CNN_MRI outperforms both the raw_MRI (PE: 3%; area under the curve (AUC) difference: 0.06) and norm_MRI (PE: 3%; AUC difference: 0.05)
The peak is in the same threshold range [3, 5] for all models, which means, recalling Figure 3, the percentage of selected voxels is between approximately 18% and 0.4%

Summary

Introduction

Technological advancements and the availability of large amounts of labeled data (Aiello et al, 2019) fostered neuroimaging research’s development (Traverso et al, 2020). In this context, machine learning (ML) algorithms played a relevant role (Lundervold and Lundervold, 2019; Chatterjee et al, 2020; Salmanpour, 2020; Traverso et al, 2020). Concerning classification tasks, some ML studies focused on support vector machines (SVMs) and random forests for binary classification of pathological versus healthy conditions ML algorithms were successfully applied to perform different tasks, like image classification, object detection, and image segmentation (Dora et al, 2017).

Objectives

Methods

Results

Conclusion