A Research On Automatic Segmentation Of Fetal Brain Tissues Based On Self-Attention Neural Networks

Congenital diseases are one of the leading causes of infant mortality.Research has shown that many congenital diseases can lead to changes in the shape or volume of the fetal brain tissue.Therefore,segmenting fetal brain tissue using fetal brain magnetic resonance imaging(MRI)can help doctors better study congenital diseases and improve their diagnosis and treatment.Currently,there are several challenges in segmenting fetal brain MRI images,including the relationship between fetal brain tissue structure and gestational age or disease status,poor image quality,unclear tissue boundaries,and manual annotation.These issues not only increase the difficulty of segmentation,but also lead to the presence of noisy labels.To achieve automatic and accurate segmentation of fetal brain tissue and address the issue of noisy labels,this thesis focuses on two aspects of research: network structure and training strategy.The specific work is as follows:(1)A neural network based on self-attention mechanism has been proposed to address the issue of insufficient global information extraction capability in traditional convolutional neural networks(CNN).This thesis introduces a global information extraction module based on self-attention mechanism,which is integrated into the CNN backbone network to combine the advantages of both.The proposed method has been experimentally validated on two fetal brain MRI datasets,Fe TA 2021 and Fe TA 2022,with Dice coefficients of 85.42% and 85.82%,respectively,outperforming existing stateof-the-art methods.(2)A multi-scale deep mutual learning strategy has been proposed along with the design of an auxiliary network to complement this strategy.To address the problem of noisy labels when learning from traditional hard labels,this thesis introduces a novel training strategy called the multi-scale deep mutual learning strategy.The strategy employs a dual-branch network,with one branch being a neural network based on selfattention mechanism,and the other branch being the auxiliary network designed in this thesis.This learning strategy enables both branches of the network to learn not only from hard labels,but also from the soft labels generated by each other,thereby acquiring additional knowledge and reducing the impact of noisy labels.Additionally,an auxiliary network is designed based on the channel attention mechanism to complement this learning strategy.In experiments,the auxiliary network achieved Dice coefficients of85.39% and 85.77% on two datasets,while the main network achieved Dice coefficients of 85.79% and 86.22% on the same datasets under the multi-scale deep mutual learning strategy,demonstrating the effectiveness of this strategy.(3)A decoupled multi-scale deep mutual learning strategy has been proposed.To further improve the segmentation performance of the model,this thesis conducts research and analysis on the proposed multi-scale deep mutual learning strategy.Through derivation of the KL divergence loss function used in the mutual learning part,it is found that it can be decoupled into a part related to the target class and a part unrelated to the target class.After decoupling,the influence of each part is analyzed and the parameters are adjusted to obtain the optimal parameter combination.In experiments,the main network under this strategy achieved Dice coefficients of 86.05% and 86.55% on two datasets,demonstrating the effectiveness of this strategy.