【作者】 王涛；

【导师】 王文剑；

【作者基本信息】 山西大学 ， 人工智能， 2025， 硕士

【摘要】 肾癌作为泌尿系统的高发疾病,其早期筛查与精准诊疗对提升患者治愈率和生存率具有重要意义。计算机断层扫描(CT)影像凭借其优异的对比度和精细的空间分辨率,在肾癌诊断中得到了广泛应用。然而,肾癌三维CT图像分割仍面临边缘像素稀疏、肿瘤尺度变化大以及形状复杂且不规则等挑战。针对上述挑战,本文研究聚焦于肾癌三维CT图像的边缘增强与小肿瘤分割,采用基于U-Net的深度学习方法捕获三维CT图像特征,并采用多特征融合策略以提升肾癌三维CT图像的分割精度。主要研究内容包括:(1)提出了基于边缘增强的选择性特征融合肾癌三维CT图像分割方法(Selective Feature Fusion Based on Edge Enhancement U-Net,EE-SFF U-Net)。该方法中设计的边缘增强模块通过融合浅层特征和互补特征以及利用三个不同方向的卷积,能够充分地挖掘肾癌三维CT图像中的边缘区域特征,强化边缘信息。选择性特征融合模块利用通道注意力机制对丰富的多层次多尺度特征进行筛选,通过深浅层特征协同作用,实现了信息的深度融合与互补。此外,设计了一种融合Generalized Dice Loss和Focal Loss的混合损失函数,通过动态权重调整机制优化训练过程,有效缓解了病变区域多尺度特性及肿瘤形态不规则性带来的挑战。该策略在确保病灶整体定位精度的同时,显著提升了对难分割目标特征的识别与提取能力。(2)针对肾癌三维CT图像中的小肿瘤,提出了一种上下文感知局部-全局特征自适应融合肾脏小肿瘤分割网络(Context-aware Adaptive Local-global Feature Fusion Renal Small Tumor Segmentation U-Net,LG-AF U-Net)。利用局部特征增强模块和全局特征学习模块来提取综合特征,增强网络对小尺度肿瘤分割的鲁棒性。自适应多尺度特征融合模块整合特征以消除冗余信息,加强特征的综合利用以解决小肿瘤区域的尺度变化。此外,基于Ki TS19数据集构建了一个小肿瘤数据集s-Ki TS19,用于评估LG-AF U-Net在处理普通方法难以分割的小肿瘤时的分割性能。实验结果表明,LG-AF U-Net显著提高了小肿瘤分割性能。在Ki TS19和Ki TS21公开数据集上的进一步实验也表明,LG-AF U-Net对于各种规模的肿瘤都能获得良好的分割结果。(3)搭建并实现了一套基于肾癌三维CT影像的智能化分割系统,旨在直观呈现所提算法的性能。该平台融合了传统CT图像处理算法与我们提出的两种新型算法,提供了从图像预处理、分割模型到三维分割的全流程解决方案。通过直观的可视化界面,用户能够便捷地进行肾癌CT图像的分析与处理操作。该平台不仅具备强大的科研功能,支持医学影像领域的前沿研究,还可作为临床辅助工具,为医生提供精准的病灶定位与量化分析支持。综上所述,为了改善肾癌三维CT图像中边缘像素稀疏、肿瘤尺度变化大和形状复杂且不规则等问题,本文设计了两种网络模型EE-SFF U-Net和LG-AF U-Net以实现了肾癌三维CT图像的高精度分割。通过在Ki TS19、Ki TS21以及s-Ki TS19小肿瘤数据集上的对比实验和消融实验,验证了所提方法的合理性和卓越性。同时,搭建的肾癌三维CT图像智能分割系统为后续的优化发展和实际应用提供了助力。更多还原

【Abstract】 Renal cancer,as a prevalent disease in the urinary system,holds significant importance in improving patient cure rates and survival rates through early screening and precise diagnosis and treatment.CT images have been widely used in the diagnosis of renal cancer because of their excellent contrast and spatial resolution.However,CT image segmentation of renal cancer still have some problems,such as sparse edge pixels,large changes in tumor scale,complex and irregular shapes and so on.To address these issues,this paper focuses on edge enhancement and small tumor segmentation in 3D CT image segmentation of renal cancer.A U-Net-based deep learning method is utilized to capture features from the 3D CT images,and the multi-feature fusion strategy is adopted to improve the segmentation performance.The research contents include:(1)A selective feature fusion based on edge enhancement U-Net(EE-SFF U-Net)is proposed.In this method,In this method,the edge enhancement module designed can fully explore the edge features in the 3D CT images of renal cancer and strengthen the edge information by fusing shallow features and complementary features and using convolutions in three different directions.The selective feature fusion module uses the channel attention mechanism to filter the rich multi-level and multi-scale features,achieving deep integration and complementarity of information through the effect of deep and shallow features.In addition,a hybrid loss function combining Generalized Dice Loss and Focal Loss is designed,which optimizes the training process through a dynamic weight adjustment mechanism,effectively mitigating the challenges posed by the multi-scale characteristics of lesion regions and the irregular shapes of tumors.This strategy ensures the overall localization accuracy of lesions while significantly enhancing the recognition and extraction capabilities of hard-to-segment target features.(2)To improve the precision of small tumors in 3D CT images of renal cancer,this paper proposes a context-aware adaptive local-global feature fusion network(LG-AF U-Net).Local feature enhancement module and global feature learning module are utilized to extract comprehensive features,thereby enhancing the network’s robustness for the segmentation of small-scale tumors.Then,the adaptive multi-scale feature fusion module integrates features to eliminate redundant information,focusing on enhancing comprehensive utilization to address scale variation in small tumor areas.In addition,we construct a small tumor dataset,s-Ki TS19,based on Ki TS19,to test LG-AF U-Net’s segmentation performance on small tumors that are difficult to segment using common methods.Experimental results indicate that our method significantly improves small tumor segmentation performance.Further experiments on the Ki TS19 and Ki TS21 public datasets also show that LG-AF U-Net outperforms other methods for tumors of various scales.(3)An intelligent segmentation platform based on renal cancer 3D CT image has been developed and implemented,aiming to demonstrate the performance of the proposed algorithms.This platform integrates traditional CT image processing algorithms with our two novel algorithms,providing a comprehensive solution from image preprocessing to segmentation models and 3D segmentation.Through an intuitive visualization interface,users can conveniently perform analysis and processing operations on renal cancer CT images.The platform not only offers robust research capabilities,supporting cutting-edge studies in the field of medical imaging,but also serves as a clinical auxiliary tool,providing doctors with precise lesion localization and quantitative analysis support.In conclusion,to solve the challenges of sparse edge pixels,large changes in tumor scale,and complex and irregular shapes in renal cancer 3D CT images,this paper designs two network models,EE-SFF U-Net and LG-AF U-Net,to achieve high-precision segmentation of 3D CT images for renal cancer.Through comparative experiments and ablation studies on the Ki TS19,Ki TS21 public datasets,and the s-Ki TS19 small tumor dataset,the rationality and superior performance of the proposed methods are validated.Additionally,the developed intelligent 3D CT image segmentation platform for renal cancer provides support for subsequent optimization development and practical applications.更多还原