Year: 2024
Type: dissertation
Abstract: Medical imaging involves the technique and process of generating visual representations of a patient's body for clinical analysis and medical intervention.Healthcare professionals heavily depend on medical images for accurate diagnosis and treatment.In clinical practice, segmentation is typically performed manually.However, when processing a vast number of medical images, the quality of segmentation can vary based on the expertise of the medical professional.This variability underscores the need for a more consistent layers.MDSU-Net uses DSC layer to reduce model complexity without degrading model performance, which is suitable for different segmentation tasks.MDSU-Net registers a Dice score of 0.7055 on ATLAS, 0.9760 on CHAOS, and 0.8883 on NERVE, respectively.Notably, these scores outperform the state-of-the-art (SoTA) benchmarks.Additionally, in terms of 3D medical image segmentation, we propose Dual Encoder U-Net (DEU-Net), which uses Transformer and CNN respectively to extract 3D medical image features in the encoder.Transformer is a pre-trained model in BTCV, which improves its ability to capture contextual features of medical images and increases the learning speed.Besides, we introduce CBAM with each convolutional layer in the encoder part to enhance CNN's feature extraction capabilities for 3D medical images.To fuse the two kinds of features, we proposed a Dual Feature Fusion Module (DFFM) to fuse the features extracted from the Transformer and CNN in the encoder, making full use of the feature extraction capabilities of the two extractors for datasets of different sizes.We compare DEU-Net with other SoTA methods on the BraTS 2020 and BraTS 2021 datasets, and the results show that the performance of DEU-Net has improved in 3D medical image segmentation task.
Author: Zhou, Yuxiang
Institution: Tokushima University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7145787761