Multimodal Integration of Brain Images for MRI-Based Diagnosis in Schizophrenia.

Raymond Salvador,Amalia Guerrero-Pedraza,Salvador Sarró,Diana Tordesillas-Gutiérrez,Peter Mckenna,Erick Canales-Rodríguez,Edith Pomarol-Clotet,Teresa Maristany,Benedicto Crespo-Facorro

doi:10.3389/fnins.2019.01203

Abstract

Magnetic resonance imaging (MRI) has been proposed as a source of information for automatic prediction of individual diagnosis in schizophrenia. Optimal integration of data from different MRI modalities is an active area of research aimed at increasing diagnostic accuracy. Based on a sample of 96 patients with schizophrenia and a matched sample of 115 healthy controls that had undergone a single multimodal MRI session, we generated individual brain maps of gray matter vbm, 1back, and 2back levels of activation (nback fMRI), maps of amplitude of low-frequency fluctuations (resting-state fMRI), and maps of weighted global brain connectivity (resting-state fMRI). Four unimodal classifiers (Ridge, Lasso, Random Forests, and Gradient boosting) were applied to these maps to evaluate their classification accuracies. Based on the assignments made by the algorithms on test individuals, we quantified the amount of predictive information shared between maps (what we call redundancy analysis). Finally, we explored the added accuracy provided by a set of multimodal strategies that included post-classification integration based on probabilities, two-step sequential integration, and voxel-level multimodal integration through one-dimensional-convolutional neural networks (1D-CNNs). All four unimodal classifiers showed the highest test accuracies with the 2back maps (80% on average) achieving a maximum of 84% with the Lasso. Redundancy levels between brain maps were generally low (overall mean redundancy score of 0.14 in a 0–1 range), indicating that each brain map contained differential predictive information. The highest multimodal accuracy was delivered by the two-step Ridge classifier (87%) followed by the Ridge maximum and mean probability classifiers (both with 85% accuracy) and by the 1D-CNN, which achieved the same accuracy as the best unimodal classifier (84%). From these results, we conclude that from all MRI modalities evaluated task-based fMRI may be the best unimodal diagnostic option in schizophrenia. Low redundancy values point to ample potential for accuracy improvements through multimodal integration, with the two-step Ridge emerging as a suitable strategy.

Highlights

In recent years, there has been growing interest in employing brain magnetic resonance imaging (MRI) datasets for medical diagnosis (Wang and Summers, 2012)
Relying on two matched samples of patients with schizophrenia (N = 96) and healthy controls (N = 115) for which structural T1, task-based, and resting-state fMRI had been acquired in a single Magnetic resonance imaging (MRI) session, we pursue three objectives: (i) to evaluate the differential discriminative power of brain maps derived from the different modalities; (ii) to quantify the degree to which the different types of images have similar or distinct predictive patterns; and (iii) to explore the added accuracy provided by a set of multimodal strategies based on different levels of data integration, including novel approaches such as a two-step data integration scheme and a one-dimensional-convolutional neural network (1D-CNN)
The highest redundancies were observed between pairs of maps that came from the same MRI modality (i.e., 1back-2back maps and amplitude of low-frequency fluctuations (ALFF)-global brain connectivity (GBC) maps) but even these rarely reached redundancy score (RSC) of 0.5 [this only occurred for RSC(GBC|ALFF) = 0.51 with Random Forests]

Summary

Introduction

There has been growing interest in employing brain magnetic resonance imaging (MRI) datasets for medical diagnosis (Wang and Summers, 2012). To improve the accuracy levels provided by unimodal data sources, some authors have explored ways to combine the information contained in images generated by different MRI modalities. These include methods with different levels of data integration and of a very different nature, ranging from simple post-classification majority-vote strategies to multimodal fusion techniques (Calhoun and Adali, 2009; Sui et al, 2013), including multiple kernel learning (Peruzzo et al, 2015; Zu et al, 2016), multimodal Gaussian process classifiers (Young et al, 2013), and deep learning (Suk et al, 2014; Shi et al, 2018) among other techniques.

Methods

Results

Conclusion