Study On Self-attentive Mechanism Algorithm For Glioma Segmentation

Glioma,a prevalent primary intracranial tumor,originates from glial cells in the brain and accounts for approximately 70% to 80% of all brain tumors.Gliomas are classified as either low-grade or high-grade,depending on the extent of invasion and the patient’s prognosis.Low-grade gliomas exhibit slower growth rates and are associated with longer survival periods.However,if left untreated,low-grade gliomas often progress into high-grade gliomas,leading to reduced survival rates and an escalating mortality risk.Therefore,the timely detection,diagnosis,and treatment of gliomas hold significant clinical importance,emphasizing the need for early intervention.MRI has emerged as a commonly employed medical imaging tool for glioma due to its ability to provide high soft tissue contrast and resolution.Accurate diagnosis of glioma heavily relies on the initial step of MRI-based segmentation.However,glioma MRI images often exhibit uneven greyscale,indistinct tumor borders,varying spatial locations,irregular morphology,and areas of high resemblance to normal brain tissues.Manual segmentation of gliomas by physicians is not only time-consuming and labor-intensive but also yields inconsistent results.Hence,it is imperative to develop a high-precision glioma segmentation algorithm to enable early clinical diagnosis and intervention.This algorithm would address the challenges posed by the inherent characteristics of glioma MRI,ensuring more efficient and reliable segmentation results.Existing MRI image segmentation algorithms for glioma face limitations such as a narrow field of view,inadequate capability to extract morphological features of glioma location,and an inability to effectively segment the fuzzy boundaries and discrete lesion areas of glioma due to a lack of attention to detailed information.In order to address these challenges in glioma image segmentation research,this project proposes two deep learning-based algorithm models.The first model introduces a segmentation algorithm based on a hybrid shifted axial self-attention mechanism.This algorithm automatically segments the tumor area using MRI images and incorporates a learnable relative position encoding that adjusts based on the significance of the relative position relationship.This enables a more accurate understanding of the global correlation between features.To tackle the issue of imprecise boundary pixel segmentation resulting from the fuzzy boundaries and unstable shapes of gliomas,a global feature fusion module is designed.This module extracts multi-scale global information and utilizes a hybrid loss function that specifically focuses on boundary pixel points.By employing multi-level axial self-attention modules,the module continuously aggregates neighboring features and learns their relative position relationships at different scales.Furthermore,a CNN-based local feature extraction module is employed to adapt to the varying morphological features of gliomas.This module extracts detailed features at multiple scales and combines them with global features to mitigate feature loss caused by downsampling.The algorithm achieves remarkable performance on the glioma test set,with a Dice coefficient of 0.8433 and a Hausdorff distance of 2.587.The second model introduces a segmentation algorithm based on a rank-variable deformable self-attention mechanism.This algorithm demonstrates the ability to accurately identify spatial correlations between features based on their significance,enabling the fusion of multi-scale information for precise learning of the morphological and structural characteristics of gliomas.To tackle the issue of uneven grayscale in glioma MRI images,a multi-scale local contrast enhancement module is designed to enhance image contrast at different scales.Additionally,a feature extraction module based on a deformable convolutional neural network is utilized to capture local deformation features.The algorithm also incorporates a ranking mechanism to emphasize important features and minimize the impact of irrelevant features.Remarkably,the algorithm achieves a Dice coefficient of 0.8695 and a Hausdorff distance of 2.199 on the glioma test set.In conclusion,this study presents two deep learning-based algorithms for brain glioma segmentation,effectively addressing challenges such as blurred boundaries,varied shapes,and non-uniform grayscale in MRI images.The proposed algorithms successfully combine the self-attention mechanism with a multi-scale feature fusion module,resulting in high accuracy in glioma segmentation,as evidenced by a Dice coefficient of 0.8433 and a Hausdorff distance of 2.587.These algorithms hold significant clinical value for early diagnosis and intervention of brain gliomas,potentially enhancing the accuracy and efficiency of glioma segmentation in MRI images.Further research could explore the application of these algorithms to other medical image segmentation tasks and evaluate their generalization capabilities.