PresentationPDF Available

Analysis of feature importance in deep neural networks in psychiatric disorders using magnetic resonance imaging



Current methods to diagnose psychiatric disorders are based on possible manifestations of the disease and behavioral criteria. Symptoms and manifestations overlap between disorders rendering the diagnostic process extremely difficult. Neuroimaging provides information about the structure and function of the brain. This information, combined with DL techniques, has a huge potential shortening and improving the diagnostic process. In this work, we train a neural network to differentiate between healthy subjects and patients of six different mental illnesses with accuracy of 64%. Finally, we analysed the network weights of the model to identify the most important regions of the brain for classification.
Analysis of feature importance in deep
neural networks in psychiatric disorders
using magnetic resonance imaging
Irina Sánchez 1, Carles Soriano-Mas 2,3, Antonio Verdejo-García 4, Narcís Cardoner 3,5,
Fernando Fernández-Aranda 2,6, José Manuel Menchón 2,3, Paulo Rodrigues 1,
Vesna Prčkovska 1 and Matt Rowe 1
2019 - Irina Sanchez - QMENTA® - 2
Speaker Name: Santi Puch
I have the following financial interest or relationship to disclose with
regard to the subject matter of this presentation:
Company Name: QMENTA
Type of Relationship: Employee and stock options holder
Declaration of
Financial Interests or Relationships
2019 - Irina Sanchez - QMENTA® - 3
No priors
Avoid bias from
prior knowledge
Subjective evaluation of the
Objective biomarker based on
anatomy or function.
Extraction of
high-level features
Deep Neural Networks
2019 - Irina Sanchez - QMENTA® - 4
of brain
Several mental disorders
- Obesity
- Anorexia
- Binge eating disorder
- Major depressive disorder
- Mild cognitive impairment
- Obsessive compulsive disorder
143 healthy subjects
137 patients with psychiatric disorders
DSK atlas rs-fMRI connectome1
1. Hagmann, P., et al.. (2008).
Mapping the structural core of human cerebral cortex.
PLoS biology, 6(7):e159.
2019 - Irina Sanchez - QMENTA® - 5
Model and Training Details
Common training details
L1 and L2 regularization.
Optimizer Adagrad with 1E-3 as learning
Early stopping based on accuracy
Batch size: 15
Dense (64)
Dense (32)
Dense (16)
50 % dropout
Dense (32)
Dense (16)
Dense (8)
60% dropout
Cross validation strategy
10 different partitions of the dataset
2019 - Irina Sanchez - QMENTA® - 6
Analysis of Feature importance
Feature vector that describes which elements of the input are most
relevant to a neuron.
Fisher’s scores to select most discriminative
Layer J+1Layer J
... W
Kim, J., et al. (2016). Deep neural network with weight sparsity control and pre-training extracts hierarchical features and enhances classification
performance: Evidence from whole-brain resting-state functional connectivity patterns of schizophrenia. Neuroimage, 124:127–146.
2019 - Irina Sanchez - QMENTA® - 7
Analysis of Feature importance
Layer 1Layer N ...
Relevant regions
Relevant regions
Relevant regions
Relevant regions
Top ranked regions per
2019 - Irina Sanchez - QMENTA® - 8
Results: Volumetry
Atrophy of
structures that
surround them
Human behaviour
Inferior frontal
Reward system
Accumbens area
Brain stem
Images obtained using BrainNet Viewer
Xia, M., Wang, J., and He, Y. (2013). Brainnet viewer: a network visualization tool for human brain connectomics. PloS one, 8(7):e68910.
Lateral Ventricles
64.04% Accuracy
62.38% F1-score
2019 - Irina Sanchez - QMENTA® - 9
Results: rs-fMRI Connectome
Memory Processing
of different
63.33% Accuracy
63.24% F1-score
Images obtained using BrainNet Viewer
Xia, M., Wang, J., and He, Y. (2013). Brainnet viewer: a network visualization tool for human
brain connectomics. PloS one, 8(7):e68910.
2019 - Irina Sanchez - QMENTA® - 10
Different findings than prior knowledge
Mix of diseases
Errors in measurements
Small database
No prior information
Highlighted areas are directly related with
psychiatric disorders present in the
Biomarker discovery framework
Applicable to other domains and tasks
Thank you for your
Santi Puch
Booth 1010
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Deep gray matter structures are associated with memory and other important functions that are impaired in Alzheimer’s disease (AD) and mild cognitive impairment (MCI). However, systematic characterization of the subregional atrophy and deformations in these structures in AD and MCI still need more investigations. In this article, we combined complex volumetry- and vertex-based analysis to investigate the pattern of subregional structural alterations in deep gray matter structures and its association with global clinical scores in AD (n = 30) and MCI patients (n = 30), compared to normal controls (NCs, n = 30). Among all seven pairs of structures, the bilateral hippocampi and nucleus accumbens showed significant atrophy in AD compared with NCs (p < 0.05). But only the subregional atrophy in the dorsal–medial part of the left hippocampus, the ventral part of right hippocampus, and the left nucleus accumbens, the posterior part of the right nucleus accumbens correlated with the worse clinical scores of MMSE and MOCA (p < 0.05). Furthermore, the medial–ventral part of right thalamus significantly shrank and correlated with clinical scores without decreasing in its whole volume (p > 0.05). In conclusion, the atrophy of these four subregions in bilateral hippocampi and nucleus accumbens was associated with cognitive impairment of patients, which might be potential target regions of treatment in AD. The surface analysis could provide additional information to volume comparison in finding the early pathological progress in deep gray matter structures.
Full-text available
Deep learning (DL) is a family of machine learning methods that has gained considerable attention in the scientific community, breaking benchmark records in areas such as speech and visual recognition. DL differs from conventional machine learning methods by virtue of its ability to learn the optimal representation from the raw data through consecutive nonlinear transformations, achieving increasingly higher levels of abstraction and complexity. Given its ability to detect abstract and complex patterns, DL has been applied in neuroimaging studies of psychiatric and neurological disorders, which are characterised by subtle and diffuse alterations. Here we introduce the underlying concepts of DL and review studies that have used this approach to classify brain-based disorders. The results of these studies indicate that DL could be a powerful tool in the current search for biomarkers of psychiatric and neurologic disease. We conclude our review by discussing the main promises and challenges of using DL to elucidate brain-based disorders, as well as possible directions for future research.
Full-text available
The human brain is a complex system whose topological organization can be represented using connectomics. Recent studies have shown that human connectomes can be constructed using various neuroimaging technologies and further characterized using sophisticated analytic strategies, such as graph theory. These methods reveal the intriguing topological architectures of human brain networks in healthy populations and explore the changes throughout normal development and aging and under various pathological conditions. However, given the huge complexity of this methodology, toolboxes for graph-based network visualization are still lacking. Here, using MATLAB with a graphical user interface (GUI), we developed a graph-theoretical network visualization toolbox, called BrainNet Viewer, to illustrate human connectomes as ball-and-stick models. Within this toolbox, several combinations of defined files with connectome information can be loaded to display different combinations of brain surface, nodes and edges
Full-text available
There is growing interest regarding the role of the right inferior frontal gyrus (RIFG) during a particular form of executive control referred to as response inhibition. However, tasks used to examine neural activity at the point of response inhibition have rarely controlled for the potentially confounding effects of attentional demand. In particular, it is unclear whether the RIFG is specifically involved in inhibitory control, or is involved more generally in the detection of salient or task relevant cues. The current fMRI study sought to clarify the role of the RIFG in executive control by holding the stimulus conditions of one of the most popular response inhibition tasks-the Stop Signal Task-constant, whilst varying the response that was required on reception of the stop signal cue. Our results reveal that the RIFG is recruited when important cues are detected, regardless of whether that detection is followed by the inhibition of a motor response, the generation of a motor response, or no external response at all.
Full-text available
This review aims to relate recent findings describing the role and neural connectivity of the basal ganglia to the clinical neuropsychiatry of basal ganglia movement disorders and to the role of basal ganglia disturbances in "psychiatric"' states. Articles relating to the relevant topics were initially collected through MEDLINE and papers relating to the clinical conditions discussed were also reviewed. The anatomy and connections of the basal ganglia indicate that these structures are important links between parts of the brain that have classically been considered to be related to emotional functioning and brain regions previously considered to have largely motor functions. The basal ganglia have a role in the development and integration of psychomotor behaviours, involving motor functions, memory and attentional mechanisms, and reward processes.
Background: Morphologic changes of the brainstem in major depressive disorder (MDD) have rarely been reported in neuroimaging studies, even though, monoaminergic neurotransmitters are synthesized in several brainstem regions. We aimed to investigate volume changes in each region of the brainstem and their association with antidepressant use or the remission status of MDD. Methods: A total of 126 patients with MDD and 101 healthy controls underwent T1-weighted structural magnetic resonance imaging. We analyzed volumes of each brainstem region, including the medulla oblongata, pons, midbrain, and superior cerebellar peduncle, and the volume of the whole brainstem using the FreeSurfer. Results: The patients with MDD had significantly greater midbrain volumes (P=0.013) compared to healthy controls. In particular, drug-naïve patients with MDD had significantly greater brainstem volumes compared to healthy controls (P=0.007), while no significant findings were observed between the antidepressant treatment group and healthy controls. The remitted patient group had reduced pons (P=0.002) and midbrain (P=0.005) volumes compared to healthy controls, while the non-remitted MDD patient group had significantly greater midbrain volumes compared to the healthy controls (P=0.017). Limitations: We could not distinguish gray versus white matter volumes changes in our analysis. Conclusions: We observed that the midbrain is enlarged in patients with a current depressive episode, who are not undergoing antidepressant treatment. This volume then returns to normal after antidepressant treatment, and is even reduced, when the patient is in remission. Further studies are needed to confirm our observations.
Apathy is characterized by lack of interest, loss of initiative, and flattening of affect. It is a frequent, very disabling nonmotor complication of Parkinson's disease (PD). The condition may notably occur when dopaminergic medications are tapered after the initiation of subthalamic stimulation and thus can be referred to as “dopaminergic apathy.” Even in the absence of tapering, some patients may develop a form of apathy as PD progresses. This form is often related to cognitive decline and does not respond to dopaminergic medications (dopa-resistant apathy). We aimed at determining whether dopa-resistant apathy in PD is related to striatofrontal morphological changes. We compared the shape of the striatum (using spherical harmonic parameterization and sampling in a three-dimensional point distribution model [SPHARM-PDM]), cortical thickness, and fractional anisotropy (using tract-based spatial statistics) in 10 consecutive patients with dopamine-refractory apathy, 10 matched nonapathetic PD patients and 10 healthy controls. Apathy in PD was associated with atrophy of the left nucleus accumbens. The SPHARM-PDM analysis highlighted (1) a positive correlation between the severity of apathy and atrophy of the left nucleus accumbens, (2) greater atrophy of the dorsolateral head of the left caudate in apathetic patients than in nonapathetic patients, and (3) greater atrophy in the bilateral nucleus accumbens in apathetic patients than in controls. There were no significant intergroup differences in cortical thickness or fractional anisotropy. Dopa-resistant apathy in PD was associated with atrophy of the left nucleus accumbens and the dorsolateral head of the left caudate. © 2014 International Parkinson and Movement Disorder Society
Background: An important step in obesity research involves identifying neurobiological underpinnings of nonfood reward processing unique to specific subgroups of obese individuals. Methods: Nineteen obese individuals seeking treatment for binge eating disorder (BED) were compared with 19 non-BED obese individuals (OB) and 19 lean control subjects (LC) while performing a monetary reward/loss task that parses anticipatory and outcome components during functional magnetic resonance imaging. Differences in regional activation were investigated in BED, OB, and LC groups during reward/loss prospect, anticipation, and notification. Results: Relative to the LC group, the OB group demonstrated increased ventral striatal and ventromedial prefrontal cortex activity during anticipatory phases. In contrast, the BED group relative to the OB group demonstrated diminished bilateral ventral striatal activity during anticipatory reward/loss processing. No differences were observed between the BED and LC groups in the ventral striatum. Conclusions: Heterogeneity exists among obese individuals with respect to the neural correlates of reward/loss processing. Neural differences in separable groups with obesity suggest that multiple, varying interventions might be important in optimizing prevention and treatment strategies for obesity.
Several previous reviews of brain imaging studies of obsessive compulsive disorder (OCD) and related conditions have been published. This article is an update on work in the field. Although, of necessity, much of the material covered here is a repetition of these previous reviews, there is a substantial body of new work published in 1992 that comprises a major augmentation of the previous literature. Most significantly, new work in functional brain imaging with positron emission tomography (PET) and single photon computed tomography (SPECT) now allows firmer conclusions to be drawn about the brain regions that mediate the symptomatic expression of OCD.