Ruopu He

Date of Award

6-6-2024

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2024

Degree Name

Master of Science (MS)

Department

Computer Science and Engineering

First Advisor

Nam Ling

Abstract

With the continuous development of deep learning theory in the field of medical images, information technology-assisted treatment methods represented by medical image segmentation technology can help doctors to quickly determine the shape and location of the lesions and improve the diagnosis efficiency of brain tumors. Based on deep learning technology, this thesis carries out related research work on MRI image segmentation. The main contents are as follows:

To begin with, acquire and prepare the brain tumor (MRI) image segmentation dataset from the official MICCAI Society website. This involves normalizing the images, cropping, and slicing, as well as scaling the data to ensure that the dataset meets the input specifications of the deep learning model. This paper presents a medical image segmentation method based on an improved Swin U-net. Initially, an atrous spatial pyramid pooling module is introduced at the end of the encoder to capture multi-scale features, allowing the model to effectively understand image at different scales and fully extract contextual information. Subsequently, in the encoder, the original blocks are replaced with residual Swin Transformer Blocks, and on the decoder side, replaced with deep residual convolution blocks. This replacement preserves the original information and alleviates the gradient vanishing issues. Lastly, an attention gate mechanism is introduced in the skip connections, enabling the model to focus more on important features within the feature map and suppress irrelevant information, thereby improving the model's segmentation accuracy.

The experimental results show that the improved segmentation model reached a validation Intersection over Union (IoU) of 89.47%, an increase of 4.36% over the Swin U-net model, demonstrating that it can effectively enhance the accuracy of image segmentation and optimize the results of the original model.

Download

Share

COinS