Deep Learning-based Multi-parameter MRI Pancreas Segmentation

Pancreatic cancer is one of the diseases that focus on current medical technology,which seriously affects the healthy life of people.Magnetic resonance image(MRI),as a kind of image with various sequence parameters,can provide doctors with detailed information of pancreatic lesions.Rapid and accurate automatic localization and segmentation of the abnormal pancreas on MRI images can help doctors find the lesion area more quickly and take effective treatment measures.However,the huge variability in the shape,size,and location of the abnormal pancreas makes automatic segmentation particularly difficult.To achieve automatic segmentation of the abnormal pancreas,this paper proposes a dual-branch interactive fusion network based on multi-scale guided feature reconstruction.The network takes advantage of magnetic resonance imaging and uses T1 Water Phase images and T1 Out of Phase images as the input of a two-branch network,which are complementary in contrast and specificity.The Selective Feature Interaction Module provided in the encoder flexibly fuses the information on the two parameter sequences to enhance the information flow in the encoder of network.In the decoder part,a Multi-scale Guided Feature Reconstruction Module is designed to reconstruct low-level information.This module eliminates the redundant organ tissue information around the pancreas,narrows the gap between low-level features and high-level features,and improves the segmentation effect of the network for small target pancreas.This paper also proposes a semi-supervised training strategy based on uncertainty estimation and morphological methods.The strategy uses uncertainty estimation to evaluate the pseudo-labels generated by the network,and makes certain corrections for pseudo-labels with high uncertainty.This strategy utilizes the continuity of 3D MRI images,and adopts morphological methods to further process pseudo-labels,so as to improve the credibility of pseudo-labels.This strategy can use unlabeled data to assist in fine-tuning the network,reducing the massive need for accurate,time-consuming manual labeling.The algorithm proposed in this paper was experimented on 395 3D MRI data of pancreatic cancer patients,104 of which were annotated with pancreatic regions.This algorithm is compared with the current advanced segmentation network.After quantitative and qualitative analysis,it can be seen that the algorithm proposed in this paper has a certain competitiveness in the segmentation of the diseased pancreas.