End-to-end tumor segmentation method based on multi-attention mechanism

Abstract

The invention relates to an end-to-end tumor segmentation method based on a multi-attention mechanism. The method mainly comprises a backbone network part and an attention module part, the backbone network comprises three sub-networks, and the three sub-networks are composed of improved 3D Resinual U-net composition; and the attention module is composed of a specially designed double-branch structure. The method can make up the defect of low training efficiency in the prior art; defect of poor segmentation precision, converting the multi-class segmentation problem of a plurality of tumor sub-regions into a plurality of two-class segmentation tasks. The attention mechanism takes the segmentation result of the tumor peripheral edema region as a soft attention and adds the soft attention intothe segmentation subtask for the tumor core part, and the segmentation result of the tumor core part is also added into the segmentation subtask for the enhancement region in the tumor core through the attention mechanism. The method is suitable for segmentation of 3D images of tumor lesion tissues with similar hierarchical structures including brain tumors, including MRI images, CT images and the like, and a more accurate segmentation result can be provided.

Description

technical field [0001] The present invention relates to the fields of computer vision, medical image processing, and image segmentation, especially the field of automatic segmentation of nuclear magnetic resonance images of tumors, and specifically relates to an end-to-end tumor segmentation method based on a multi-attention mechanism. Background technique [0002] The precise division and description of different tissues of tumors is a very important step in treating tumors and judging the prognosis. In current clinical practice, tumor delineation is usually done by professional doctors or experienced radiologists. However, medical images are different from traditional 2D images in natural scenes. Many tumors need to be diagnosed with 3D images generated by MR, CT and other instruments. Manual segmentation of these 3D images is often time-consuming, labor-intensive and costly. In addition, different annotators have different judgment standards and annotation principles for...

AI Technical Summary

Problems solved by technology

[0004] In view of the complexity of the task, many current works solve the segmentation of multiple tumor sub-regions by designing a more powerful model, but these single models are less effective for the segmentation of smaller regions in the image

There is also a lot of work by decomposing complex tasks, using multiple simple models to process different subtasks, and finally combining them through cascading, but one disadvantage of this cascading network is that it is not an end-to-end network, so its training and testing efficiency is low. At the same time, the overall task is decomposed. Although each subtask can achieve a local optimum, it cannot achieve a global optimum.

Images