Method for training model of deep neural network, and method and system using the model to determine whether there is abnormality in brain

A method for training a model of a deep neural network (DNN) that is to be used to determine whether there is an abnormality in a brain, the method comprising steps of:

• receiving a plurality of reference three-dimensional (3D) brain images including 3D images of normal brains and 3D images of brains of patients having neuropsychosis, each of the plurality of reference 3D brain images having a plurality of voxels that compose a multi-voxel pattern;
• for each of the plurality of reference 3D brain images, implementing a calibration procedure on the reference 3D brain image by
• obtaining, from the reference 3D brain image, three reference sectional images that are taken respectively on three mutually-orthogonal anatomical planes,
• determining whether each of the reference sectional images meets a corresponding given geometric condition by inputting the three reference sectional images respectively into three orientation-determining DNN models, and
• when it is determined that one of the reference sectional images does not meet the corresponding given geometric condition, rotating the reference 3D brain image about an axis perpendicular to one of the anatomical planes, on which said one of the reference sectional images is taken, so that the reference 3D brain image is in an orientation where said one of the reference sectional images meets the corresponding given geometric condition;
• for each of the plurality of reference 3D brain images that has been processed by the calibration procedure, generating a first image of white matter (WM), a second image of gray matter (GM) and a third image of cerebrospinal fluid (CSF) based on the multi-voxel pattern of the reference 3D brain image;
• and
• inputting the first images of WF generated respectively for the plurality of reference 3D brain images into a first DNN model, the second images of GM generated respectively for the plurality of reference 3D brain images into a second DNN model, and the third images of CSF generated respectively for the plurality of reference 3D brain images into a third DNN model to train the first, second and third DNN models,
• wherein the first, second and the third DNN models that have been trained are able to determine whether there is an abnormality in the white matter, whether there is an abnormality in the gray matter and whether there is an abnormality in the cerebrospinal fluid, respectively.