摘要
目的早期肾癌可以通过肾肿瘤剜除术进行有效治疗,为了降低手术难度和减少手术并发症,需要对手术的难度进行合理有效的评估。本文将深度学习、医学影像组学和图像分析技术进行结合,提出一种基于CT(computed tomography)影像的肾肿瘤剜除术难度自动评估方法。方法首先建立一个级联的端到端分割模型对肾脏、肾肿瘤和腹壁同时进行分割,同时融入子像素卷积与注意力机制,保证了小体积肿瘤分割的精确性;然后使用影像组学特征对误判的肾肿瘤进行去除;最后依据分割结果,采用国际标准的梅奥肾周粘连概率(Mayo adhesive probability,MAP)评分和R.E.N.A.L评分对肾脏和肾肿瘤进行自动化的评估计算,并根据计算结果得出肾肿瘤剜除术难度。结果将实验的自动化评估结果与三甲医院泌尿科的3位医疗专家的结果进行对比,从预测的平均结果来看,超过两个专家,与最好的专家相差仅0.1%。平均预测时间,单个肿瘤约为244 ms,标准差只有8 ms,专家评估时间约为26 s,标准差在3 s左右,自动评估速度是人工的108倍左右。结论自动化评估结果整体上与专家评估水平基本一致,同时评估速度更加快速稳定,可以有效替代专家进行自动化评估,为术前准确诊断、手术方案个体化规划和手术入路选择提供准确可靠的决策支持,给手术难度诊断评估提供智能化的医疗解决方案。
Objective Early renal cancer can be identified and treated effectively via enucleation of renal tumor.To opti‐mize surgery and its surgical complications,it is necessary to evaluate the surgical feasibility efficiently and effectively.To quantify the difficulty index of surgical contexts,Mayo adhesive probability(MAP)score and R.E.N.A.L score can be involved in for its applications.computed tomography(CT)images-based manual analysis is roughly estimated in terms of these two scoring standards-related corresponding difficulty score.To optimize the accuracy and reliability of evaluation,this sort of qualitative manual evaluation method is time-consuming and labor-intensive.Thanks to deep learning technique based medical radiomics and image analysis,we develop an automatic evaluation method of CT images-based surgical optimization in relevance to enucleation of renal tumors.Method First,a three-layer cascade end-to-end segmentation model is illustrated to segment the kidney,renal tumor and abdominal wall simutanesoualy.Each layer is linked with an extended U-Net for segmentation.The abdominal wall segmentation is at the top of them,followed by the kidney segmentation,and the renal tumor is at the bottom.This stratification is derived of a learning process of spatial constraints.For the extended U-Net,the dense connection is reflected in the convolution block of the coding layer or coding and decoding layer-between same layer,as well as upper and lower-between layers.This kind of dense connection at the three levels can be used to obtain more semantic connections and transmit more information in the training,and overall gradient flow can be effectively enhanced,and the global optimal solution can be sorted out smoothly.To alleviate the loss of redundant texture detail in the up-sampling process,the sub-pixel convolution mode is used further.This method proposed can generate higher resolution images through the pixel order-related intervention of multiple low resolution feature images.At the same time,image mode-med
作者
刘云鹏
吴铁林
蔡文立
王仁芳
孙德超
干开丰
李瑾
金冉
邱虹
徐惠霞
Liu Yunpeng;Wu Tielin;Cai Wenli;Wang Renfang;Sun Dechao;Gan Kaifeng;Li Jin;Jin Ran;Qiu Hong;Xu Huixia(International Exchange College,Ningbo University of Technology,Ningbo 315100,China;Li Huili Hospital Affiliated to Ningbo University,Ningbo 315000,China;Radiology Imaging Laboratory,Harvard Medical School,Boston 02114,USA;Zhejiang Wanli University,Ningbo 315000,China)
出处
《中国图象图形学报》
CSCD
北大核心
2023年第8期2461-2475,共15页
Journal of Image and Graphics
基金
国家自然科学基金项目(61906170)
浙江省基础公益研究计划项目(LQ21H060002,LGF21F020023,LGF21F020022,LGF19F020008)
宁波市自然科学基金项目(2019A610033,202003N4324)
宁波市科技局一般项目(2019C50008)
宁波市重大科技公关项目(20211ZDYF020177)
宁波市公益类科技计划项目(2021S105)。
关键词
肾肿瘤剜除术
医学图像分割
影像组学
深度学习
手术评估
enucleation of renal tumor
medical image segmentation
radiomics
deep learning
surgical evaluation
