Systems and methods for diagnosing autism spectrum disorder using fMRI data

A method for diagnosing autism spectrum disorder (ASD), the method comprising:

• receiving, by a processor, functional magnetic resonance imaging (fMRI) data comprising a plurality of brain scans obtained from a plurality of patients, respectively;
• using, by the processor, exactly one sparse autoencoder (SAE) on the fMRI data to extract testing patterns and generate a testing feature vector, the testing feature vector being generated using only the exactly one SAE;
• providing, by the processor, the testing feature vector as a testing input to a deep neural network (DNN);
• and
• running, by the processor, the DNN on the testing feature vector to determine which brain scans of the plurality of brain scans are indicative of ASD, said brain scans that are indicative of ASD thereby resulting in a diagnosis of ASD in the corresponding patients, of the plurality of patients, from which said brain scans were obtained.

A system for diagnosing autism spectrum disorder (ASD), the system comprising:

• a processor;
• and
• a machine-readable medium in operable communication with the processor and having instructions stored thereon that, when executed by the processor, perform the following steps:
• receiving functional magnetic resonance imaging (fMRI) data comprising a plurality of brain scans obtained from a plurality of patients, respectively;
• using exactly one sparse autoencoder (SAE) on the fMRI data to extract testing patterns and generate a testing feature vector, the testing feature vector being generated using only the exactly one SAE;
• providing the testing feature vector as a testing input to a deep neural network (DNN);
• and
• running the DNN on the testing feature vector to determine which brain scans of the plurality of brain scans are indicative of ASD, said brain scans that are indicative of ASD thereby resulting in a diagnosis of ASD in the corresponding patients, of the plurality of patients, from which said brain scans were obtained,
• the DNN comprising an input layer, two hidden layers, and a softmax output layer,
• the instructions when executed further performing the following steps:
• training the exactly one SAE prior to using the exactly one SAE to extract testing patterns and generate the testing feature vector;
• and
• training the DNN prior to running the DNN on the testing feature vector,
• the training of the exactly one SAE and the training of the DNN being performed simultaneously,
• the training of the exactly one SAE and the training of the DNN comprising running a predetermined number of iterations of training on the exactly one SAE and the DNN, each iteration comprising:
• i) extracting time series from the fMRI data;
• ii) calculating Pearson's correlations on the time series;
• iii) selecting ¼ smallest and ¼ largest average pairwise correlations of the Pearson's correlations to generate a training feature vector and a SAE loss;
• iv) providing the training feature vector as a training input to the DNN;
• v) running the DNN on the training feature vector to generate a prediction loss;
• vi) combining the SAE loss and the prediction loss to generate a total loss;
• vii) providing the total loss to the exactly one SAE and the DNN;
• and
• viii) returning to step i),
• the predetermined number of iterations being at least 30,
• the exactly one SAE comprising a cost function comprising a mean squared error, a weight decay, and a sparsity term, and
• the training of the exactly one SAE comprising minimizing the cost function of the exactly one SAE.