计算机科学
分割
人工智能
图像分割
图像(数学)
任务(项目管理)
深度学习
相关性(法律)
计算机视觉
尺度空间分割
卷积神经网络
机器学习
模式识别(心理学)
基于分割的对象分类
医学影像学
点(几何)
作者
Abdullah Shazly,Abbas Mohammed,Abdulkhalek Al-Fakih,Mohammed A. Al‐masni
标识
DOI:10.1109/idap68205.2025.11222376
摘要
Medical image segmentation remains a critical yet challenging task, requiring high precision in delineating complex anatomical structures and pathological regions. In this study, we present a systematic evaluation of three representative deep learning paradigms for 3D medical image segmentation: the Optimized U-Net (CNN-based), Swin UNETR (transformer-based), and Diff-UNet (diffusion-based). These models are benchmarked across two widely used datasets, called BraTS 2023 (brain MRI) and the Medical Segmentation Decathlon Task 08 (liver CT), under a unified experimental framework. Our results show that, despite architectural innovations, transformer- and diffusionbased models do not consistently outperform CNN-based methods. The Optimized U-Net achieves the highest segmentation accuracy across both datasets, highlighting the enduring strength and clinical relevance of well-tuned convolutional networks. These findings present the importance of rigorous empirical validation and suggest that architectural complexity alone does not guarantee superior performance in medical image analysis.
科研通智能强力驱动
Strongly Powered by AbleSci AI