Research On Abnormal Signal Detection Method Of Focal Cortical Dysplasia Based On Multimodality And Deep Learning

Focal Cortical Dysplasia(FCD)is the main cause of Drug-Refractory Epilepsy(DRE).Surgical resection of epileptic lesions is an important choice for the treatment of FCD.However,manual evaluation of FCD lesions is time-consuming and laborious,and an automatic detection method for FCD is urgently needed in clinic.However,at present,the amount of FCD data is very limited.How to realize FCD detection through a spot of data is a huge technical problem.On the other hand,the FCD lesions are not clear,the labels are not accurate,and the model cannot be correctly fitted.How to identify the lesions through the low precision labels is a challenge to overcome the automatic detection of FCD.In the first part,this thesis realizes the preliminary detection of FCD lesions through multimodal data and multimodal fusion coding.First of all,this thesis improves UNet for multimodal data,and proposes a multimodal fusion encoding network(Multi-Encoder UNet,ME-UNet),which uses multimodal independent encoding and long connection structure to effectively extract features from multimodal data.Then,using the morphological analysis method,the key morphological features are extracted.At the same time,the introduction of PET data enables the model to learn the pathogenesis of metabolic abnormalities.Finally,ME-UNet is used to fit the multimodal data to realize the preliminary segmentation of FCD lesions.In the second part,this thesis improves ME-UNet and designs a hybrid lateral connection structure to fuse shallow detail features with deep features.Furthermore,this thesis introduces a self-attention mechanism to the deep features of the encoder,supplements the channel features,and further improves the performance of the model.Finally,based on the above structure,this thesis proposes a multi-channel network based on hybrid lateral connection(Multi-Encoder with Hybrid Lateral Connection Net,MEHLC-Net).In this thesis,the average Dice index obtained through the ME-UNet network and multimodal data can reach 0.42,which has reached the clinical detection requirements.After subsequent improvements,the MEHLC-Net index can further reach 0.48,which proves the advanced nature of the network.