Pyramid pooling enhanced ResUNet for accurate 3D brain image segmentation

Abstract

"Medical picture segmentation is important for clinical applications because it can offer valuable information on disease identification. With the inclusion of deep learning techniques, the original U-Net and ResUnet architecture has demonstrated excellent performance in 2D medical picture segmentation issues. However, it is still challenging to extend the U-Net to handle 3D volumetric medical images. This thesis proposed a redesigned ResUnet architecture with a hybrid model called pyramid pooling with enhanced ResUNet fusion with ResUnet and dialated spatial pyramid pooling from DeepLabV3+. Therefore, CNNs will effectively aid us in addressing the 3D segmentation problem.

Accurate segmentation of 3D brain pictures is critical in neuroimaging research because it allows for exact anatomical localization and quantitative analysis. In this paper, we introduce a novel framework for robust and high-fidelity 3D brain image segmentation that combines the capability of Dilated Spatial Pyramid Pooling (DSPP) with the Residual U-Net (ResUNet) architecture. The ResUNet’s DSPP module improves multi-scale feature representation by aggregating information across several spatial resolutions, allowing the network which represent feature context. This integration tackles the issues given by complicated brain architecture as well as the unpredictability in picture quality that is frequent in real-world datasets in a synergistic manner. The model can comprehend complex patterns and recognize minute details in medical images thanks to attention processes, residual connections, and feature fusion methods. Brain tumors are divided in the research into medical images where the clinical data or benchmark datasets will be used to assess the proposed model. In order to assess segmentation accuracy and contrast it with cutting-edge techniques, The Dice similarity coefficient metrics will be used.

This paper will create a novel and efficient 3D image segmentation framework using a modified ResUNet architecture and enhanced pyramid pooling from DeeplabV3+."