Denoising Task-Based Functional Magnetic Resonance Imaging Data Using Hybrid Deep Neural Network Model




Kashfolayat, Sahar

Journal Title

Journal ISSN

Volume Title



Error management is one of the most important concerns in medical science. A doctor's ability to make an accurate diagnosis can save a patient's life. Therefore, it is essential that doctors have access to accurate medical data in order to make the best possible decisions for their patients. This means that technology can help doctors be more accurate in their diagnoses and reduce the number of mistakes they make. In this study, a hybrid deep neural network (HDNN) is proposed to reduce the noise in task-based functional magnetic resonance imaging (fMRI) data. This model has been developed to improve the performance of the Deep Neural Network (DNN) model [1]. This network, like the DNN model, has a sequential structure and utilizes the same cost function and the same algorithm to update network parameters. The architecture of the model consists of three common layers and four non-common layers to perform noise reduction. The first three layers consist of two one-dimensional convolutional filters to reduce physiological noise in both GM and non-GM data and one LSTM layer to consider the temporal correlation of input data [9]. The final layers are used to separate and select data respectively, including a fully connected and a conventional selective layer for GM and non-GM input data (with different resolutions) which introduces the hybrid operation of the network based on the type of input data. In order to analyze the effect of these modifications on the output image, using the same conditions as described in [1], simulated and real fMRI data are used to compare DNN and HDNN models. The results show that the proposed model was able to reduce noise better than the DNN model by at least 5% in GM areas and 83% in non-GM areas.



fMRI noise reduction