Description
Autism spectrum disorder (ASD) is a complex neurological and developmental disorder. It emerges early in life and is generally associated with lifelong disability. Finding the biomarkers associated with ASD is extremely helpful to understand the underlying roots of the disorder and find more targeted treatment. Previous studies suggested brain activations are abnormal in ASDs, hence functional magnetic resonance imaging (fMRI) has been used to identify ASD. In this work we addressed the problem of interpreting reliable biomarkers in classifying ASD vs. control; therefore, we proposed a 2-step pipeline: 1) classifying ASD and control fMRI images by deep neural network, and 2) finding which brain regions are important for identifying ASD and control. Specifically, in step 2, we used the trained classifier to estimate the feature importance by measuring the prediction distribution change as a function of input image with the corrupted region. However, there is no certain way to corrupt the data without adding side effects. Thus, we aggregated two "opposite" corruption methods: a) blackout and b) add Gaussian noise. Biomarkers found by the 2-step pipeline were verified by Neurosynth brain function decoding. Several key innovations in our research include: i) we created an innovative pipeline for learning image data feature by analyzing the classifier outcomes with corruptions; ii) we proposed a deep learning strategy for classifying 4D data; iii) we aggregated different corruption methods for feature importance analysis, and iv) our neurological interpretation of the final results showed evidence that there were meaningful fMRI biomakers on fMRI for ASD.
ASD Biomarker Detection on fMRI Images: Feature learning with Data Corruptions by Analyzing Deep Neural Network Classifier Outcomes
Autism spectrum disorder (ASD) is a complex neurological and developmental disorder. It emerges early in life and is generally associated with lifelong disability. Finding the biomarkers associated with ASD is extremely helpful to understand the underlying roots of the disorder and find more targeted treatment. Previous studies suggested brain activations are abnormal in ASDs, hence functional magnetic resonance imaging (fMRI) has been used to identify ASD. In this work we addressed the problem of interpreting reliable biomarkers in classifying ASD vs. control; therefore, we proposed a 2-step pipeline: 1) classifying ASD and control fMRI images by deep neural network, and 2) finding which brain regions are important for identifying ASD and control. Specifically, in step 2, we used the trained classifier to estimate the feature importance by measuring the prediction distribution change as a function of input image with the corrupted region. However, there is no certain way to corrupt the data without adding side effects. Thus, we aggregated two "opposite" corruption methods: a) blackout and b) add Gaussian noise. Biomarkers found by the 2-step pipeline were verified by Neurosynth brain function decoding. Several key innovations in our research include: i) we created an innovative pipeline for learning image data feature by analyzing the classifier outcomes with corruptions; ii) we proposed a deep learning strategy for classifying 4D data; iii) we aggregated different corruption methods for feature importance analysis, and iv) our neurological interpretation of the final results showed evidence that there were meaningful fMRI biomakers on fMRI for ASD.