Lesion Segmentation Based On Pseudo-healthy Synthesis

In recent years,deep learning algorithms represented by deep convolutional neural networks have achieved remarkable results in the fields of image recognition,detection,and segmentation.In the task of segmentation of medical imaging lesions,the fully convolutional neural network represented by U-Net has achieved close or even surpasses in the brain,liver,heart,kidney and other parts as well as ultrasound images,magnetic resonance images and other forms of images.The performance of human experts is still limited by the problem of "big data,small tasks"(high requirements for training and test data,and narrow applicable tasks)and cannot be well applied to actual clinical scenarios.This paper explores the health knowledge existing in health images and pixels,and embeds it in the traditional supervised paradigm neural network by generating pseudo health images corresponding to the lesion images,so as to improve the generalization,robustness and reliability of the network.Explanatory.Specifically,under different data conditions,this article respectively proposes the following two lesion segmentation schemes.For the pseudo-health image generation and lesion segmentation tasks under the condition of having the lesion image and its segmentation label,this paper proposes a pseudohealth image generation framework that confronts the generation task and the segmentation task,and enhances the generation of images by "deceiving" the continuously strengthened segmenter.Visual health.At the same time,improved residual loss and weighted cross-entropy loss for this task are proposed to further improve the visual quality of the generated image.Finally,the obtained pseudo-health image is embedded into the segmentation model to solve the problem of low contrast and blurred boundary of the lesion.A large number of experiments on two public data sets(BraTS and LiTS)verify the effectiveness of the proposed scheme and multiple modules.At the same time,it is also proved through experiments that embedding pseudo-health images into the segmentation model can improve the segmentation performance.For pseudo-health image generation and unsupervised abnormal segmentation tasks under the condition of scarcity of images containing lesions,this paper uses the rich multimodal information in medical images.Specifically,due to the difference in the imaging process,the abnormal signal contrast of different modalities is different.In this paper,we use weak-contrast modal images to generate pseudo-health images corresponding to strong-contrast modal images,and detect the lesion by comparing the difference between the generated image and the real strong-contrast modal image.In the experimental stage,this paper verifies the effectiveness of the proposed model on two public data sets(BraTS and ISLES).