噪声标注场景下的医学图像分割OA
Clean-Label-Guided Boundary Denoising Distillation for Medical Image Segmentation under Noisy Annotation
针对医学图像分割中精细标注成本高、粗边界标注易误导模型学习的问题,提出少量精细标注引导的边界去噪蒸馏方法(CL-BDD).该方法以U-Net为基础网络,将训练样本划分为精细标注集和噪声标注集,并采用多边形边界退化模拟快速轮廓标注中的边界简化与局部偏移.教师模型联合两类标注训练,引入可靠边界约束和样本形态分布;学生模型使用教师生成的软伪标签替代粗糙硬标签,使边界区域保留概率过渡,减弱错误轮廓监督的影响.在JSRT和BUS_UC数据集上的实验表明,CL-BDD在BUS_UC测试集上取得Dice 0.876 0、BF1 0.554 4,BF1较直接使用噪声标签训练提高0.155 7;在JSRT测试集上取得Dice 0.994 9、BF1 0.940 6.消融实验表明,少量精细标注和软伪标签均有助于改善边界质量.结果表明,少量可靠标注可为粗边界监督提供校正信号,在不精确噪声标注下改善分割模型的边界质量.
To address the high cost of fine annotation and the tendency of coarse boundary labels to mislead model learning in medical image segmentation,this paper proposes a clean-label-guided boundary denoising distillation method(CL-BDD).The method uses U-Net as the backbone network,divides the training samples into a fine-annotation set and a noisy-annotation set,and applies polygonal boundary degradation to simulate boundary simplification and local displacement in rapid contour annotation.A teacher model is trained jointly with both types of annotations to introduce reliable boundary constraints and sample's morphological distribution,while a student model uses the soft pseudo-labels generated by the teacher to replace coarse hard labels.In this way,probabilistic transitions are preserved in boundary regions,and the influence of erroneous contour supervision is reduced.Experiments on the JSRT and BUS_UC datasets show that CL-BDD achieves a Dice of 0.876 0 and a BF1 of 0.554 4 on the BUS_UC test set,im-proving BF1 by 0.155 7 compared with direct training using noisy labels.On the JSRT test set,CL-BDD achieves a Dice of 0.994 9 and a BF1 of 0.940 6.Ablation experiments demonstrate that both a small number of fine annotations and soft pseudo-labels contribute to improved boundary quality.These results indicate that effectively leveraging a small number of fine annotations can improve segmen-tation performance under imprecise noisy boundary supervision.
丁熠;林玉茹;夏煜超
电子科技大学信息与软件工程学院,四川成都 611731||电子科技大学长三角研究院(衢州),浙江衢州 324003电子科技大学信息与软件工程学院,四川成都 611731电子科技大学信息与软件工程学院,四川成都 611731
信息技术与安全科学
医学图像分割噪声标注边界去噪蒸馏多边形边界退化软伪标签
medical image segmentationnoisy annotationboundary denoising distillationpolygonal boundary degradationsoft pseudo-label
《四川师范大学学报(自然科学版)》 2026 (6)
727-737,11
国家自然科学基金(62576079、62372083)、四川省自然科学基金(2026NSFSC0415)、四川省科技计划项目(2024NSFTD0005)、成都医学院2024年第一批临床科学研究基金重点项目(24LHXX1-01、24LHXX1-02)、四川省卫生健康委员会2023年科学技术项目——临床研究专项基金(23LCYJ011)、残障人士智能康复及无障碍教育部重点实验室开放基金项目(2024KFJJ004)
评论