Self-supervised tumor segmentation system based on layer decomposition

Abstract

The invention provides a self-supervised tumor segmentation system based on layer decomposition. The self-supervised tumor segmentation system comprises: a module M1, which acquires a texture generation simulated tumor area from a normal image through a randomly generated three-dimensional mask and transformation; a module M2, which is used for fusing the simulated tumor area with the normal image to obtain a synthesized tumor image; and a module M3, which is used for training a deep convolutional neural network to learn layer decomposition, acquiring a tumor segmentation image from the synthesized tumor image by using the trained deep convolutional neural network, recovering a tumor region image, recovering the normal image and synthesizing an image. According to the method, the tumor image can be synthesized according to the normal image, and through an effective layer decomposition-based self-supervised learning model, the features with relatively strong discrimination capability on the tumor area are extracted, so tumor segmentation under the condition of no label is realized.

Description

technical field [0001] The invention relates to the technical fields of computer vision and image processing, in particular to a self-supervised tumor segmentation system based on layer decomposition. Background technique [0002] In recent years, with the rapid development of computer vision, existing deep neural network-based algorithms have achieved considerable progress in various image analysis tasks. For medical image data, compared with natural image data, there are more modalities, and usually 3D images, which pose great challenges in data processing and analysis. Medical image segmentation is the most basic step in medical image analysis. More and more algorithms based on deep learning have been proposed, which can effectively provide various intermediate representations such as position, shape and size to assist doctors in disease diagnosis. However, existing medical image segmentation algorithms usually rely on a large amount of labeled data. However, due to the ...

AI Technical Summary

Problems solved by technology

However, the models trained by these pre-defined pseudo-label tasks usually cannot be directly applied to downstream target tasks, and certain labeled data are still required to fine-tune the model

Images