| Intestinal tract is the main part of human digestive system,which has the functions of digestion,absorption and movement.Colorectal cancer is a malignant tumor with high morbidity and mortality worldwide,and can grow anywhere in the colon or rectum.However,most cases of colorectal cancer grow in the form of adenomatous polyps in the intestinal lining in the early stage.Although the growth rate is slow,the painless and painless characteristics in the early stage still lead patients to ignore the health screening for colorectal polyps.Therefore,most cases of colorectal cancer are already in the middle and advanced stage or even worse when detected.This seriously threatens the overall prognosis and survival rate of patients.At present,gastroscopy is still the gold standard for screening gastrointestinal polyps.The gastroscopy images collected by penetrating into the digestive tract can help us clearly and accurately observe the information of the morphology and tissue scope of polyps.However,it is still a challenging task to accurately segment polyps on the acquisition of images,because polyps are highly similar to the intestinal lining where they grow,and their shapes are different.Meanwhile,the impact of intestinal peristalsis and foreign bodies greatly reduces the accuracy of polyp segmentation.In this paper,the convolutional neural network framework based on deep learning is used to segment digestive tract polyp images.According to the image characteristics and clinical application requirements,three kinds of digestive tract polyp segmentation algorithms are designed.The main contents are as follows:(1)A new colorectal polyp segmentation model based on double codec architecture was proposed.The segmentation prediction generated by the upstream model is used as the attention weight chart to provide gated attention guidance to the downstream model to strengthen the target subject recognition ability of the downstream model.In order to further improve the utilization rate of the upstream output weight,a multi-scale pooling attention mechanism was proposed.The translation invariance brought by the convolutional collocation pooling operation was used to provide multi-scale attention weight graphs at different resolutions for the downstream model,which further improved the segmentation effect of the model and refined the segmentation edge.The pre-trained Res2 Net is introduced as the Backbone to further enhance the feature extraction ability of the model.The model in this chapter mainly solves the problem of the segmentation effect of the current algorithm in the field of colorectal polyp segmentation,and conducts research experiments for the purpose of improving the overall segmentation index.(2)A lightweight codec architecture for colorectal polyp segmentation model based on improved Ghost Net was proposed.By using cavity convolution and maximum pooling to improve the specific module in Ghost Net V2,a lightweight coding module(DCG-bneck)which is more suitable for segmentation tasks can be obtained and used as the backbone network.At the same time,in order to optimize the common channel redundancy problem of feature matrix in the process of jump connection and decoding,a subspace attention mechanism based on feature graph grouping was proposed.The large channel feature matrix was divided into multiple sub-matrix sequences,and the coordination attention(CA)was calculated for them respectively.In this way,the multi-dimensional attention calculation could be obtained while the feature redundancy was used more efficiently.Improve model segmentation efficiency;Finally,in order to further reduce the influence of high feature redundancy on the decoding process,a decoding module combining channel mixing and depth-separable convolution is proposed.In addition to reducing the amount of computation,the feature matrix in the decoding process is grouped,and the probability of feature redundancy is reduced by random exchange of feature maps between different groups.The algorithm in this chapter is a lightweight model,which mainly solves the problems of high storage,high complexity and low efficiency of large models in the field of colorectal polyp segmentation.(3)In order to combine the advantages of the large model and the lightweight model proposed above,a structured knowledge distillation architecture is proposed.The large model proposed in Chapter 3 is used as the teacher model to guide the student model in Chapter 4.By adding Sigmoid activation function of distillation temperature to soften the teacher’s prediction label,KL dispersion is used to generate pixel-level evaluation of the student model prediction distribution,and the calculation of pixel-level knowledge distillation loss based on soft label is realized.At the same time,in order to utilize the feature extraction capability of the large model,a feature fusion distillation method under multi-scale is proposed to whiten the multi-scale feature matrix separated from the downstream decoding structure of the teacher model,so as to realize the separation and normalization of trainable parameters.Then,the weighted losses of BCEWith Logits Loss and Smooth L1 Loss were calculated with the decoding feature matrix of the corresponding resolution in the student model.The final feature distillation loss was defined as the mean loss of the feature matrix at the resolution of 16,32,64,128 and 256.The Ground Truth of the input data is equally important.The hard label loss is also a part of the total loss in this chapter.Finally,the weighted sum of these three losses is taken as the total loss of the structured knowledge distillation framework in this chapter.The parameters of the teacher model are directly obtained through pre-training,and no parameter updating of the teacher model is involved in the whole training stage.The parameters of the student model are updated according to the gradient descent method based on loss function.The overall segmentation performance of the student model can be further improved without increasing the number of parameters,and the model compression based on knowledge transfer can be realized. |
