【作者】 刘博；

【导师】 周柚；

【作者基本信息】 吉林大学 ， 工程硕士（专业学位）， 2025， 硕士

【摘要】 肺癌是全球范围内发病率和死亡率最高的恶性肿瘤之一,肺结节作为其早期表现,其检测与随访的准确性对患者预后具有重要影响。肺部作为运动器官,在呼吸过程中发生大范围形变,肺结节体积小。传统方法难以应对这种形变,导致配准精度低,检测过程中依赖人工判读,效率低且主观性强。近年来,深度学习技术在医学图像分析领域取得了显著进展,但针对肺CT的配准与肺结节检测仍面临处理大范围形变时表现不佳、小目标检测难、缺少差异性量化标准等挑战。针对上述问题,本文开展了评价肺结节差异性计算方法的研究,该方法包含配准、检测和差异性评价三个阶段。首先,本文构建了一种基于特征金字塔的配准网络(Feature Pyramid Displacement field Prediction Attention Network,FPANet),有效缓解了大范围形对配准精度的影响。随后,本文进一步提出了一种轻量化小目标敏感的肺结节检测网络(YOLO Attention Network,YOLO-Attn Net),通过标记目标框计算得出肺结节的数量、位置和大小。最后,本文设计并实现一种多时相肺结节差异化计算方法(Multi-temporal Lung Nodule Differential Calculation Method,MLNDCM),计算出同一个体在不同时期肺结节的差异性。本文的主要工作内容如下:1、本文提出了配准网络FPANet以解决具有大形变特质的肺CT图像配准问题。FPANet首先通过特征金字塔网络(Feature Pyramid Network,FPN)实现多尺度特征提取,随后利用本文提出的耦合了U-Net和注意力机制的位移场预测网络(Prediction Attention Network,PANet)增强对关键区域关注,来提升配准精度。此外,本文在相似性度量的基础上,引入KL散度和位移场平滑损失,来提高FPANet在图像形变较大的动态场景下的配准精度。实验表明,相较于当前最优深度学习方法,FPANet平均配准误差至少降低了4%,位移场平滑性得到显著改善。2、本文提出了一种轻量化小目标敏感的肺结节检测网络YOLO-Attn Net。针对肺结节尺寸较小、边界不清晰的问题,YOLO-Attn Net提出了小目标增强注意力机制模块(Detail Scale Coord Attention,DSCA),替换掉了YOLOv11网络结构中的后半部分,增强了模型对小目标的感知能力。此外,本文提出了由交互比损失、尺度自适应权重和边界敏感项组成的小目标增强损失(Small Target Io U Loss,STIOU),来提高YOLO-Attn Net对小目标的敏感性和对目标框边界的感知能力。实验结果表明,YOLO-Attn Net对小尺寸肺结节的检测精度达到79.5,与YOLOv11相比平均精度提升约5%,模型参数量降低了30%,表现出较强的鲁棒性和准确性。3、本文提出了一种多时相肺结节差异化计算方法MLNDCM。为量化评价同一个体在不同时期肺结节变化的,MLNDCM利用YOLO-Attn Net的检测结果计算出不同时期肺结节之间的位移、大小及长宽比等差异,以直观反映肺结节的差异变化。更多还原

【Abstract】 Lung cancer is one of the most prevalent and fatal malignancies worldwide.As an early manifestation of lung cancer,the detection and follow-up accuracy of pulmonary nodules have a significant impact on patient prognosis.The lungs,as a dynamic organ,undergo substantial deformation during the breathing process,while pulmonary nodules are typically small in size.Traditional methods struggle to address such deformations,leading to low registration accuracy and reliance on manual interpretation,which is inefficient and subjective.In recent years,deep learning technologies have made significant strides in the field of medical image analysis.However,challenges remain in the registration and detection of pulmonary CT images,particularly in handling large deformations,detecting small targets,and lacking a standardized method for quantifying differences.To address these issues,this paper investigates a method for evaluating the differentiation of pulmonary nodules,which encompasses three stages:registration,detection,and differentiation evaluation.First,a feature pyramid-based registration network(Feature Pyramid Displacement Field Prediction Attention Network,FPANet)is proposed,which effectively alleviates the impact of large deformations on registration accuracy.Next,we introduce a lightweight,small-target-sensitive lung nodule detection network(YOLO Attention Network,YOLO-Attn Net),which calculates the number,location,and size of pulmonary nodules by labeling bounding boxes.Finally,we design and implement a multi-temporal lung nodule differentiation calculation method(Multi-temporal Lung Nodule Differential Calculation Method,MLNDCM),which computes the differences in pulmonary nodules at different time points for the same individual.The main contributions of this paper are as follows:(1)This paper proposes FPANet to solve the challenge of registering lung CT images with large deformations.FPANet first extracts multi-scale features through a Feature Pyramid Network(FPN),and then enhances attention to key regions by incorporating a displacement field prediction network(Prediction Attention Network,PANet)that combines U-Net and attention mechanisms.This improves registration accuracy.Additionally,we introduce KL divergence and deformation field smoothing loss on top of similarity measures to enhance FPANet’s registration performance in dynamically deformed scenarios.Experimental results demonstrate that FPANet reduces the average registration error by at least 4%compared to the current state-of-the-art deep learning methods,with a significant improvement in the smoothness of the deformation field.(2)This paper proposes a lightweight,small-target-sensitive lung nodule detection network,YOLO-Attn Net,to address issues with the small size and unclear boundaries of pulmonary nodules.YOLO-Attn Net introduces a small-target-enhancing attention mechanism module(Detail Scale Coord Attention,DSCA),replacing the latter part of the YOLOv11 network architecture to improve the model’s sensitivity to small targets.Furthermore,we introduce a small-target-enhancing loss(Small Target Io U Loss,STIOU),consisting of interaction ratio loss,scale-adaptive weights,and boundary-sensitive terms,to enhance YOLO-Attn Net’s sensitivity to small targets and its ability to perceive target boundary details.Experimental results show that YOLO-Attn Net achieves a detection accuracy of 79.5%for small pulmonary nodules,with an approximate 5%average precision improvement over YOLOv11,while reducing model parameters by 30%,demonstrating strong robustness and accuracy.(3)This paper proposes the MLNDCM to quantify and evaluate the changes in pulmonary nodules over time for the same individual.This method utilizes the detection results from YOLO-Attn Net to compute the displacement,size,and aspect ratio differences of pulmonary nodules across different time points,providing an intuitive reflection of nodule changes over time.更多还原