Publicação
Mapping of resting-state networks in resting-state functional. Magnetic resonance with deep learning
| Resumo: | Pre-surgical mapping relies on neuroimaging techniques, such as Functional Magnetic Resonance (fMRI). Resting-State Functional Magnetic Resonance Imaging (rs-fMRI) enables the detection of brain spontaneous activity which have formed coherent Resting-State Networks (RSNs). Multiple techniques have been used to map the representation of function using rsfMRI, among these are Deep Learning (DL) methods. Based on current literature, we studied the feasibility of implementing a 3D Convolutional Neural Network for mapping RSNs. To create and evaluate our model, we used the Autism Brain Imaging Data Exchange I (ABIDE-I) dataset. An additional dataset from Hospital da Luz (HL) was used in testing stages. After data and Regions of Interest (ROIs) selection, brain maps of correlation were obtained and assigned to their respective RSNs. Subsequently, the main core architecture of our model was defined. In order to achieve better classification performance results, several trials were required. These varied according to the number of filters or neurons presented in each convolutional and fully connected layers, as well as the set of hyperparameters implemented. After performing model training, it was time to test our model. Together with accuracy results, AUC values and confusion matrices were obtained, allowing a better insight of the predictions for all classes. In addition, several model evaluation measures, such as precision, sensitivity and F1-score were acquired. All the results gathered allowed an overall comparison between the two datasets implemented at this stage. The results of this dissertation revealed that the proposed pipeline was able to classify RSNs. Overall, it demonstrated the feasibility of implementing a 3D CNN to map several RSNs, elucidating the relevance of both rs-fMRI and DL in mapping brain regions. |
|---|---|
| Autores principais: | Azevedo, Carolina Isabel Santos |
| Assunto: | Ressonância Magnética Funcional de Repouso Redes de repouso Aprendizagem profunda Redes Neuronais Convolucionais Tese de mestrado - 2023 |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | Pre-surgical mapping relies on neuroimaging techniques, such as Functional Magnetic Resonance (fMRI). Resting-State Functional Magnetic Resonance Imaging (rs-fMRI) enables the detection of brain spontaneous activity which have formed coherent Resting-State Networks (RSNs). Multiple techniques have been used to map the representation of function using rsfMRI, among these are Deep Learning (DL) methods. Based on current literature, we studied the feasibility of implementing a 3D Convolutional Neural Network for mapping RSNs. To create and evaluate our model, we used the Autism Brain Imaging Data Exchange I (ABIDE-I) dataset. An additional dataset from Hospital da Luz (HL) was used in testing stages. After data and Regions of Interest (ROIs) selection, brain maps of correlation were obtained and assigned to their respective RSNs. Subsequently, the main core architecture of our model was defined. In order to achieve better classification performance results, several trials were required. These varied according to the number of filters or neurons presented in each convolutional and fully connected layers, as well as the set of hyperparameters implemented. After performing model training, it was time to test our model. Together with accuracy results, AUC values and confusion matrices were obtained, allowing a better insight of the predictions for all classes. In addition, several model evaluation measures, such as precision, sensitivity and F1-score were acquired. All the results gathered allowed an overall comparison between the two datasets implemented at this stage. The results of this dissertation revealed that the proposed pipeline was able to classify RSNs. Overall, it demonstrated the feasibility of implementing a 3D CNN to map several RSNs, elucidating the relevance of both rs-fMRI and DL in mapping brain regions. |
|---|