Brain tumor detection method based on attention mechanism and MRI (Magnetic Resonance Imaging) multi-modal fusion

Abstract

The invention discloses a brain tumor detection method based on attention mechanism and MRI (Magnetic Resonance Imaging) multi-modal fusion, and the method comprises the steps: replacing a common convolution block of an encoder with a mixed cavity convolution block based on a Multi-Unet model; a multi-branch output convolution block MB-OutConv for short is autonomously designed by referring to a multi-branch encoder structure of an Inception model; designing an attention module CB-Attention based on channels, capturing pixel point association among the channels of the original segmented image, and performing attention weighting on the channels; a neural network is properly improved, a new attention module is independently designed to further perfect a segmentation result, and the attention module is based on an image channel and completes attention weighting at a pixel point level. And finally, segmenting the tumor and other lesion areas in the brain MRI image. On the basis of a multi-modal convolutional neural network Multi-Unet, part of encoder branches are improved, and an attention module is added behind the multi-modal convolutional neural network Multi-Unet to jointly improve the segmentation effect of the brain tumor.

Description

technical field [0001] The invention relates to deep learning, the field of medical image processing, and is an interdisciplinary field of computers and medicine. Based on the multi-modal convolutional neural network, combined with the current challenges, the neural network is appropriately improved, and a new attention module is independently designed to further improve its segmentation results. Attention weighting is done at the point level. The ultimate goal is to segment tumors and other diseased regions in brain MRI images. Background technique [0002] Image segmentation is the process of covering the target contour with a pixel-level mask. Since the shape of the lesion area is irregular, the doctor must judge the location of the disease as accurately as possible when diagnosing, otherwise, the healthy tissue will be damaged. Surgery is performed as a diseased area, which will put the patient's life in danger, so image segmentation is more meaningful than target dete...

AI Technical Summary

Problems solved by technology

The 3D convolution kernel itself has a relatively large amount of parameters, and this is only the number of one channel. If it is used to extract the source 3D MRI image and store the features in multiple channels, the overall data volume will be very large. Memory and video memory will not be able to support

The existing segmentation network, regardless of whether the model is based on 2D or 3D, can segment the general outline of the lesion area. In addition to the need to consider the connection between 2D images, there are still some areas for improvement in terms of convolution, such as in T1 , In the T1c image, most of the contours are not very clear, especially the edema area, and this phenomenon also leads to the problem of excessive segmentation when the network extracts features.

Images