| dc.description.abstract | Clinical statistics reveal an average survival period of approximately fifteen months for malignant intracranial tumors, with an almost 100% recurrence rate within two years. Early assessment, diagnosis, and treatment are crucial in addressing this challenging prognosis. Therefore, the success of automated brain tumor recognition and segmentation tasks in medical imaging technology plays a vital role in assisting physicians with diagnosis and guiding patient treatment strategies. Manual annotation of tumor locations in magnetic resonance imaging (MRI) images and medical professionals planning surgical pathways is time-consuming. To improve the accuracy of tumor localization and contour marking while alleviating the burden on healthcare providers, this study employs geometric deep learning models for automatic brain
tumor image recognition, segmentation, and three-dimensional tumor volume reconstruction. In our research, we will use the BraTS 2020 dataset, which first undergoes fast data density functional transformation to enhance tumor features, and
combines it with Squeeze-and-excitation, which can enhance the features of three-microbrain tumors. The module is built into the Encoder-Decoder Structure of D-Unet(Dimension Fusion U-Net) to complete and train a series of geometric deep learning
models. In addition to using D-Unet, we will also use other types of deep learning models, such as nnU-Net, 3D-U-net and other contemporary most popular brain tumor segmentation models to compare the model computational complexity and Dice
segmentation score results, and then the model with the highest Dice score is improved to obtain the best segmentation results. Our research has found that the dataset after fast
data density functional transformation can significantly shorten the training and inference time of the deep learning model (about 50% or more), and can also improve the segmentation performance of the model. | en_US |