Brain-Inspired Model And Its Application To Intelligent Healthcare

With the rapid development of artificial intelligence(AI),healthcare has greatly progressed as AI can do what humans do and even surpass the performance of humans in the areas such as early detection of disease,diagnosis and assisted-living.Motivated by the success of AI,this thesis investigates a few brain-inspired AI models,which are directly inspired by how human brains process data,to address applications in the area of intelligent healthcare.The main novelties of this thesis are summarized as follows:1.A new method is proposed for detecting salient objects from a set of images,known as co-saliency detection.In this method,co-saliency detection is considered as a two-stage saliency propagation task,where saliency values in image pairs are propagated based on color similarities.The co-salient foreground cue and background cue are considered together for generating contrast-enhanced co-saliency maps.The proposed co-saliency detection method is then applied for enhancing the human-activity areas and improving the performance of human fall detection.2.Limited by visual percepts from existing visual prostheses,it is necessary to enhance their functionality to fulfill some challenging tasks for the blind such as obstacle avoidance.A spiking neural network model is first introduced for obstacle recognition by modeling and classifying spatio-temporal video data.Obstacle-related features are captured from low-resolution prosthetic vision images,followed by an integrated spiking neural network model consisting of spiking trains encoding,input variable mapping,unsupervised reservoir training and supervised classifier training.The efficiency of spike-based computation makes it possible to directly utilize available neuromorphic hardware chips,embedded in visual prostheses,to improve current prostheses.3.A novel multi-stream multi-scale convolutional network scheme is proposed to generate multi-resolution features for Alzheimer's disease detection from MRI images.The proposed scheme employs several parallel3 D multi-scale convolutional networks,each applying to individual tissue regions(GM,WM and CSF)followed by feature fusions.The proposed fusion is applied in two separate levels: the first level fusion is applied on different scales within the same tissue region,and the second level is on different tissue regions.To further reduce the dimensions of features and mitigate overfitting,a feature selection method based on XGBoost is utilized before the classification.4.To tackle the commonly encountered problems of insufficiently large brain tumor datasets and incomplete modality of image for deep learning,we propose to add augmented brain MR images to enlarge the training dataset by employing a pairwise Generative Adversarial Network(GAN)model.The pairwise GAN is able to generate synthetic MRIs across different modalities.A two-stage course-to-fine training strategy is proposed to learn the glioma feature using GAN-augmented MRIs followed by real MRIs.To achieve the3 D scan-based diagnostic result,we employ a post-processing strategy to combine the slice-level glioma classification results by majority voting.5.Currently,many available glioma datasets often contain some unlabeled brain scans,and many datasets are moderate in size.We propose to exploit deep semi-supervised learning to make full use of the unlabeled data.Deep CNN features were incorporated into a new graph-based semisupervised learning framework for learning the labels of the unlabeled data,where a new 3D-2D consistent constraint is added to make consistent classifications for the 2D slices from the same 3D brain scan.A deep-learning classifier is then trained to classify different glioma types using both labeled and unlabeled data with estimated labels.