Multi-modal nuclear magnetic image cerebral arterial thrombosis lesion segmentation method based on convolutional neural network

Abstract

The invention discloses a multi-mode nuclear magnetic image cerebral arterial thrombosis lesion segmentation method based on a convolutional neural network. The method comprises the steps of designinga deep convolutional neural network 1 based on an original framework of a 3D U-Net network; basic modules based on a 3 * 3 * 3 convolution layer and a batch standardization layer; a 3D deformable convolution module based on a 3 * 3 * 3 convolution layer and trilinear interpolation; a cascade deformation module based on the 3D deformable convolution module and the basic module; constructing a deepconvolutional neural network 2 based on the 3D U-Net network and the cascaded deformation module; inputting the processed data into a neural network 1 for pre-training; assigning a weight obtained bypre-training to a neural network 2 for training; and verifying the segmentation effect on the test set of the pixel-level label. The invention provides an automatic labeling method for ischemic stroke lesion segmentation, so that the cost of labeling data is greatly reduced, the engineering operability is enhanced to a certain extent, and doctors are assisted in clinical diagnosis of ischemic stroke patients.

Description

technical field [0001] The invention relates to a convolutional neural network-based segmentation method for ischemic stroke lesions in multi-modal MRI images aimed at the clinical diagnosis of ischemic stroke patients. Background technique [0002] Ischemic stroke is the lack of oxygen to the brain tissue caused by the disorder of blood supply to the brain, and the gradual necrosis of brain cells due to the interruption of blood flow in the next few hours is too severe or the interruption is too long, and finally irreversible damage to the infarction is formed. core. [0003] According to the latest epidemiological studies, the age-standardized prevalence, annual incidence and death rate of stroke in China are 1114.8 / 100,000, 246.8 / 100,000 and 114.8 / 100,000 respectively. The government imposes a heavy economic burden. Early diagnosis and treatment of ischemic stroke can significantly improve the prognosis of patients. Magnetic resonance examination is very important for i...

AI Technical Summary

Problems solved by technology

[0005] For ischemic stroke lesions whose location, shape, and size are highly random, the U-Net network commonly used in the field of medical imaging cannot better meet the accuracy requirements of segmentation, and methods such as extended Inception Net have little improvement and introduce a lot The amount of parameters, the gradient disappearance or redundant calculation that may be caused by deepening the network, etc., the present invention proposes to combine the 3D variable convolution module with U-Net, which can adapt to the variability of the shape of the lesion, avoid the gradient disappearance of the network and efficient Feature utilization, simple network implementation

Images