Liver Tumor Image Segmentation And MVI Prediction Based On Feature Learning

Liver cancer is a malignant disease threatening human life and health.Microvascular invasion(MVI)is currently considered to be an independent risk factor for liver cancer recurrence.The preoperative prediction of MVI determines the formulation of patient treatment plan.Using magnetic resonance image instead of surgical materials to predict MVI in patients with liver cancer can not only reduce the workload of doctors,but also improve the efficiency of medical auxiliary diagnosis,which has important clinical significance.The purpose of feature learning technology is to transform the spatial distribution of original data into a more understandable expression related to tasks.Feature learning technology for medical aided diagnosis is to represent and model diseases through various types of data sets to realize intelligent diagnosis of diseases.Different from natural images,tumor heterogeneity makes medical data diversified.In order to improve the accuracy and efficiency of medical assisted diagnosis,feature learning technology should consider diversified data types,so as to build a more accurate and effective diagnosis model.Feature-based learning technology is mainly divided into domain knowledge driven and datadriven.Domain knowledge driven method is mainly based on the professional knowledge of relevant fields,through manual feature extraction and algorithm design,which has good interpretability and robustness.The disadvantage is that it needs manual participation,low efficiency and difficult model processing.The data-driven method is represented by deep learning technology.It automatically learns features through a large number of data samples.It has high efficiency,but poor interpretability and robustness,and seriously depends on the number of labeled samples.In view of the shortcomings of knowledge-driven and data-driven feature learning technology,this paper deeply studies the feature learning technology in medical aided diagnosis,especially the intelligent diagnosis method based on the combination of knowledge-driven and data-driven,which can effectively characterize diversified data types and improve the level of medical aided intelligent diagnosis.The preoperative prediction process of MVI based on magnetic resonance image includes several key application links,such as tumor segmentation,feature extraction,feature fusion and model construction.Therefore,this paper focuses on four aspects: tumor image segmentation,knowledge driven feature learning,data-driven feature learning and feature learning based on fusion strategy Automatic intelligent diagnosis from feature extraction to MVI preoperative prediction improves the ability of feature learning for diversified data types,and provides some new methods and perspectives for MVI preoperative prediction to solve practical clinical problems.Some specific research results have been obtained:1、A liver tumor segmentation algorithm based on full convolution neural network based on attention mechanism is proposed.Most of the U-shaped segmentation networks based on encoder decoder use layer hopping connection to cascade the high-level and low-level features,but they are simply combined without considering the correlation between the high-level and low-level features,resulting in the decline of segmentation accuracy.To solve this problem,this paper improves the FC densenet network,adds a global average pooled attention module(GAM)between the upper sampling channel and the lower sampling channel,uses the gam module to redistribute the weight of the high-level feature map,and uses these weight information to guide the low-level features to recover the image detail information at high resolution,It effectively uses the characteristics that high-level features focus on classification and low-level features focus on detail information recovery,forming an effective fusion of high-level and low-level feature information.The experimental results show that FC densenet network based on attention module gam can segment liver tumor MRI images effectively.Compared with other mainstream convolution neural network segmentation algorithms,it improves the Dice coefficient,average pixel accuracy and average intersection ratio by about 3%,2% and 3%.2、A feature learning method of liver tumor based on domain knowledge driven is proposed.The complexity of MVI determines that the diagnosis of MVI needs a lot of domain expert knowledge of image data and non image data.Aiming at the fusion of diversified medical data types,this paper proposes a domain knowledge driven learning method for liver tumor representation.For the segmented T1 WI and T2 WI tumor images,the artificial design features such as gray,shape,texture and wavelet that can evaluate the tumor heterogeneity are extracted.The sparse minimum optimization solution is obtained from the multi-dimensional features by using the least absolute shrinkage and selection operator lasso algorithm with L1 regularization and cross validation method,The multi collinearity problem caused by multi sequence image feature fusion is solved,and the unstructured clinical characterization indexes are integrated into the classifier,which improves the representation ability of liver tumor of diversified medical data and has good interpretability.The experimental results show that the MVI prediction results with AUC of 0.812 are obtained based on the domain knowledge driven model.3、A data-driven fine-grained feature learning method is proposed.In view of the fine-grained difference of liver tumor MVI in magnetic resonance images and the insufficient number of labeled image samples,a semi supervised dual channel network DST net based on contrast learning and channel weighting is proposed to learn the fine-grained features of magnetic resonance images.Through the training of a large number of data samples,the triple network extracts fine-grained intra class differences and inter class differences through unsupervised sample comparison learning.Channel weighted simple se densenet network extracts high-level semantic features.The dual channel network DST net integrates supervised high-level semantic features and fine-grained features,which effectively improves the representation ability of fine-grained features.The experimental results show that the semi supervised dual channel network DST net can effectively express the fine-grained features of the image,and the AUC of MVI prediction is 0.777.4、An end-to-end network MVI prediction algorithm based on multi-level feature fusion is proposed.The data-driven convolutional neural network is highly dependent on data,and while extracting high-level semantic features,it will lead to the loss of low-level information.The knowledge-driven model fully absorbs the knowledge of domain experts,and can better describe the image details from the artificially designed low-level information such as texture and wavelet,The fusion based on data-driven and knowledge-driven will further improve the feature learning ability of liver tumors.This paper proposes an end-to-end network of multi-level feature fusion,integrates the high-level semantic features,fine-grained features and domain knowledge extracted by DST net network,and constructs the feature system of global local domain knowledge features.The experimental results show that the end-to-end network of multi-level feature fusion proposed in this paper achieves an AUC of 0.826.5.A new liver tumor segmentation algorithm based on cavity convolution with stacked tree aggregation structure is proposed.In the multi-scale target segmentation task,hole convolution can extract multi-scale information and rich context information,but there is a chessboard artifact problem in hole convolution,resulting in the decline of segmentation accuracy.To solve this problem,this paper first proposes the residual dense connection module RDB in the encoder network,adds the hole convolution module tasd of tree aggregation structure to the encoder decoder network,uses the hole convolution module tasd to expand the receptive field and extract rich multi-scale information,and the tree aggregation structure can effectively eliminate the chessboard artifact caused by hole convolution,The segmentation accuracy is improved.The experimental results show that the RDB module proposed in this paper can improve the performance by 4%,and the tasd module can improve the performance by 5%.