Research On Segmentation Method Of 3D Cardiac Medical Image Based On Knowledge Distillation

Medical image segmentation is a very basic and important task in the medical field,which helps to aid diagnosis and clinical research.With the vigorous development of deep learning,the Convolutional Neural Network(CNN)method has performed very well in the field of biomedical image segmentation.However,there are two main problems in deep learning when dealing with medical image tasks: 1)Medical image data processing is difficult,the modalities of medical images are diverse,and there are often artifacts,noise,etc.,and different from natural images,medical images are difficult to obtain.resulting in a smaller amount of data.2)The deeper the depth of the convolutional neural network,the better the performance of the model is often obtained,but if the depth is too deep,it is easy to cause overfitting,and the amount of calculation and storage is large.These problems increase the difficulty of processing medical image problems based on deep learning.To overcome the above problems,researchers proposed a 3D convolutional neural network that can directly process 3D medical data.The input and output of the network are both 3D data.However,methods such as model compression and transfer learning solve the problem of huge neural network models.Among them,knowledge distillation as a model compression method has attracted the attention of researchers and has been successfully applied to tasks such as object detection and image classification.The knowledge distillation method in deep learning usually uses two different network structures as the teacher network and the student network.The teacher network chooses a more complex network,the student network chooses a lightweight network,and the student network uses the knowledge distillation loss to approximate the teacher network.The knowledge of the teacher network is transferred to the student network.This paper proposes a method for segmentation of 3D medical images based on knowledge distillation,which combines two knowledge distillation methods to improve the segmentation performance of the student network.First,this paper uses a3 D convolutional neural network and its lightweight form as a teacher network and a student network,respectively,to directly process 3D medical data.On the basis of the segmentation loss used in the conventional distillation algorithm,two distillation losses are introduced to realize the training of the end-to-end student network model,which are the prediction label distillation loss and the 3D structured distillation loss.The prediction label loss supervises the student network to learn the prediction ability explicitly from the segmentation map output by the teacher network.At the same time,according to the sequential characteristics of MRI images,a distillation method based on 3D structure is adopted to transfer the implicit knowledge in the teacher network to the student network.The trained student network is then self-distilled,and different distillation loss methods are studied to further improve the performance of the student network.Finally,a number of ablation experiments were carried out on the data set HVSMR2016 for the method proposed in this paper.It was verified that the Dice coefficient-based evaluation index has a certain improvement compared with the original student network,and compared with other distillation methods.The performance is good.