Research On Glioma Segmentation Algorithm For Missing MRI Modalities

Glioma originates from glial cells in the brain and is one of the most common types of primary brain tumors.Glioma is highly malignant and requires timely diagnosis and treatment.Generally,the automatic segmentation of glioma based on magnetic resonance images is a crucial step in treating glioma diseases.It is an important reference for assisting doctors in clinical diagnosis and surgical planning.With the rapid development of deep learning technology,more and more automatic glioma segmentation algorithms have been proposed.Since multiple MRI image modalities can provide complementary information about human brain organization,most segmentation algorithms fuse multiple MRI image modalities for training models.In theory,the joint use of multiple magnetic resonance image modalities can improve the accuracy of the glioma segmentation algorithm.However,in both single-center and multi-center scenarios,there may be a problem of missing magnetic resonance imaging modalities,which hinders the improvement of the model’s performance.Specifically,in a single-center scenario,factors such as damaged or restricted scanners and patients allergic to certain contrast agents lead to the loss of some modalities during inference,affecting the segmentation performance of the model.In a multi-center scenario,due to data privacy issues,each center may not be able to collect a complete set of multi-modality data for training,thus failing to improve the model’s performance.To address the above problems,a glioma segmentation algorithm based on generalized knowledge distillation and a glioma segmentation algorithm based on federated learning is proposed.The former utilizes multi-modality data during the training period to deal with the problem of missing modalities during inference,while the latter proposes a new federated learning training strategy to improve the model performance for each center in a multi-center scenario.The details of the proposed work are listed as follows:1)Aiming at the missing MRI modality problem in single-center scenarios,this thesis focuses on the most challenging case where only one modality can be used during inference and proposes a novel generalized knowledge distillation with a ”teacherstudent” architecture framework,which transfers the additional modality knowledge from the multimodal teacher model to the unimodal student model through two knowledge distillation strategies,namely the segmentation graph distillation strategy and the cascaded region attention distillation strategy,to improve the segmentation accuracy of the unimodal student model.Specifically,the segmentation graph distillation strategy encourages the student model to imitate the softened output of the teacher model,while the cascaded region attention distillation encourages the student model to explore and learn the regional feature information of the teacher model on multiple layers.2)Aiming at the missing MRI modality problem in multi-center scenarios,this thesis introduces a common federated learning framework used in multi-center scenarios to avoid data privacy and security issues.At the same time,a new federated learning strategy is proposed that uses personalized layers to address the problem of different centers having different numbers of modalities and introduces mutual learning distillation strategy to bring additional modal knowledge to each center to assist in model training and improve model performance.3)To explore the engineering value of the proposed algorithm,this thesis implemented a federated learning system for glioma segmentation.The system has two major functions: federated training and image segmentation.Federated training focuses on simulating the federated learning training process for glioma segmentation across multiple centers,meeting the daily research needs of medical researchers with limited experimental conditions.Image segmentation integrates the proposed algorithm model to assist doctors in daily segmentation and improve their diagnostic efficiency.This thesis presents a large number of experiments to validate the proposed algorithm,demonstrating its effectiveness and robustness.Additionally,relevant functional tests are conducted on the designed system,confirming its completeness of functionality.