Research On Semi-Supervised Learning For Segmentation Of Liver Tumor Without Contrast Agents

For accurate liver cancer diagnosis and staging,contrast-enhanced magnetic resonance imaging is one of the most commonly used modern diagnostic techniques.Although liver tumors on the enhanced images are significantly enhanced after the use of contrast agents,which is beneficial to the segmentation of liver tumors,the use of contrast agents also brings problems such as gadolinium deposition,time-consuming,and the risk of adverse reactions.Although direct segmentation on non-enhanced images can avoid these problems,direct segmentation from non-enhanced images is quite challenging due to the low contrast of non-enhanced images,let alone how to obtain accurate manual annotations from physicians in the first place.In recent years several researches has been done to address this issue,but these methods usually require fully annotated datasets,and have higher requirements for data preprocessing,data annotation and other pre-processing work.Medical datasets are generally considered small-scale datasets,often comes with few samples and labeling difficulty.Therefore,how to make the most of a medical dataset is an urgent problem to be solved.In order to directly use unlabeled data for training and reduce the reliance on manual labeling,this paper focuses on semi-supervised learning method to tackle this situation,and also combines the characteristics of multi-modal and multi-phase images of liver MRI data,which led to a series of innovative studies conducted on semi-supervised liver tumor segmentation from non-enhanced MRI data.The main research results are as follows:(1)Contrast enhanced image have the advantages of high sensitivity and high contrast between liver tumors and background,so it is an intuitive idea to use enhanced images to guide non-enhanced images.In this paper,direct guidance on augmented and non-augmented images is addressed for the first time,and a semi-supervised learning method is established by introducing consistency loss,therefore the liver tumor segmentation performance from non-enhanced images is improved.In this paper,a multi-period consistent semi-supervised model MPC-MT based on Mean Teacher is proposed,which can establish consistency regularization between multi-phase MRI images,thereby establishing a semi-supervised framework to tackle liver tumor segmentation from non-enhanced images.For experiments MPC-MT was applied to a real-world dataset consisting of 194 cases.The results show that MPC-MT achieves better results than the baseline network under different dataset annotation rates,and is comparable to the baseline network under 100% dataset annotation rate with only70% dataset annotation rate.MPC-MT achieved a Dice coefficient of 89.87% and a precision of 92.29% on the 100% annotated dataset.(2)The process to obtain contrast enhanced images will not only cause contrast agent deposition,but also have the disadvantages of time consuming and high price.Therefore,how to directly use non-enhanced images for model training is a valuable problem.In this paper,a new approach to use multiple student models on non-enhanced multi-modal data to segment each modality separately,then introduce a stability loss to achieve dynamic guidance between multiple student models is addressed.This paper proposed a stability-aware multi-student model SA-MS,which uses multiple student models to achieve segmentation on multiple modalities respectively,then generates final prediction through a multi-modal fusion module.For experiments SA-MS is applied to a 3D MRI dataset consisting of 179 patients,and the results show that the SA-MS achieves a Dice coefficient of 82.16% using only 50% of the annotated dataset,and a Dice coefficient of 90.51% can be achieved on the 100% annotated dataset.