The existing prototype learning approaches typically compute a single global prototype per category, which often struggles to achieve precise segmentation in adhesive edge regions and ambiguous areas. To address this issue, this paper proposes a context-aware prototype-based semi-supervised medical image segmentation framework. It employs a self-distillation mechanism to learn affinity relationships between local regions of annotated images and query image patches, and utilizes clustering algorithms to generate fine-grained multi-prototype representations, significantly enhancing the discriminative capability and feature granularity of prototype learning. To further improve model performance, a memory bank mechanism is incorporated into both the self-distillation and prototype generation modules, effectively leveraging reliable information from annotated images to simultaneously enhance global contextual understanding and local detail segmentation accuracy. Moreover, contrastive learning is integrated into the training framework to achieve intra-class feature compactness and inter-class feature discriminability. Extensive experiments on two widely used datasets, LA and NIH Pancreas, demonstrate that the proposed method outperforms recent semi-supervised approaches. Under the 10% and 20% labeled data settings for semi-supervised segmentation tasks, our method achieves Dice scores of 90.18% / 91.15% on the LA dataset and 79.42% / 82.25% on the NIH Pancreas dataset, respectively, surpassing state-of-the-art semi-supervised methods.
半监督医学图像分割任务使用少部分标注图像和大量无标签图像训练网络,因此训练数据集被分为标注数据和无标签数据两部分.基于均值教师-学生网络框架[4],本文提出了基于上下文感知原型合成的半监督医学图像分割框架,如图1所示.方法采用一个编码器-解码器结构的分割模型F(·)作为骨干网络.将图像输入该网络后,产生的像素级特征分别输入一个自蒸馏编码器Esd(·)和两个卷积头G(·)、H(·)中.Esd(·)与分割网络编码器的结构相同,用于后续的自蒸馏学习,目标是对当前图像及数据集中的相似区域进行分组.Esd(·)与F(·)间存在梯度截断,目的是防止自蒸馏过程影响网络中的其他部分,其输出结合自蒸馏构建图像块间的亲和度矩阵,辅助上下文感知原型合成.G(·)作为预测头,根据模型的输出特征来预测像素级分割预测概率图. H (·)是一个卷积头,其输出的像素级特征用于后续基于原型的对比学习. 图1中下标s表示学生模型,下标t表示教师模型.
对于输入图像q,对其分别执行随机空间变换Tt和Ts后得到对应变换图像,将二者分别输入学生模型和教师模型,得到教师模型自蒸馏特征及学生模型自蒸馏特征.之后通过对教师模型自蒸馏特征进行空间逆变换并应用学生网络的空间变换,确保教师模型与学生模型的自蒸馏特征间空间位置上的一致性. 由于F(·)输出的像素级特征进入自蒸馏编码器后,经过4次下采样,所以 Fsd中的每个向量 fm 都代表了q中相应位置的图像块的特征,其中m为空间位置索引.
RONNEBERGERO, FISCHERP, BROXT. U-net:convolutional networks for biomedical image segmentation[M]//Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015.Cham:Springer International Publishing,2015:234-241.
[2]
MILLETARIF, NAVABN, AHMADIS A. V-net:fully convolutional neural networks for volumetric medical image segmentation[C]//2016 Fourth International Conference on 3D Vision (3DV). October 25-28,2016. Stanford,CA,USA.IEEE,2016:565-571.
[3]
CHENB Z, LIUY S, ZHANGZ,et al .TransAttUnet:multi-level attention-guided U-Net with transformer for medical image segmentation[J]. IEEE Transactions on Emerging Topics in Computational Intelligence,2024,8(1): 55-68.
[4]
TARVAINENA, VALPOLAH. Mean teachers are better role models:weight-averaged consistency targets improve semi-supervised deep learning results[C]//Neural Information Processing Systems (Neur IPs), 2017.
[5]
YUL Q, WANGS J, LIX M,et al .Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation[C]//Medical Image Computing and Computer Assisted Intervention-ICCAI 2019.Cham:Springer,2019:605-613.
[6]
BAIY H, CHEND W, LIQ L,et al .Bidirectional copy-paste for semi-supervised medical image segmentation[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 17-24,2023,Vancouver,BC,Canada. IEEE,2023:11514-11524.
[7]
CHEND W, BAIY H, SHENW,et al .MagicNet:semi-supervised multi-organ segmentation via magic-cube partition and recovery[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 17-24,2023,Vancouver,BC,Canada. IEEE,2023:23869-23878.
[8]
LUOX D, WANGG T, LIAOW J,et al .Semi-supervised medical image segmentation via uncertainty rectified pyramid consistency[J].Medical Image Analysis,2022,80:102517.
[9]
WUY C, GEZ Y, ZHANGD H,et al .Mutual consistency learning for semi-supervised medical image segmentation[J]. Medical Image Analysis,2022,81:102530.
[10]
LUOX D, CHENJ N, SONGT, et al. Semi-supervised medical image segmentation through dual-task consistency[J]. Proceedings of the AAAI Conference on Artificial Intelligence,2021,35(10): 8801-8809.
[11]
GAOS B, ZHANGZ J, MAJ C,et al. Correlation-aware mutual learning for semi-supervised medical image segmentation[C]//Medical Image Computing and Computer Assisted Intervention-MICCAI 2023. Cham: Springer, 2023: 98-108.
[12]
SNELLJ, SWERSKYK, ZEMELR S. Prototypical networks for few-shot learning[C]//Neural Information Processing Systems,2017:4080-4090.
[13]
ZHANGX L, WEIY C, YANGY,et al .SG-one:similarity guidance network for one-shot semantic segmentation[J].IEEE Transactions on Cybernetics,2020,50(9):3855-3865.
[14]
WANGZ J, LUOY D, HUANGZ,et al .Prototype-matching graph network for heterogeneous domain adaptation[C]//Proceedings of the 28th ACM International Conference on Multimedia.Seattle, WA, USA. ACM,2020:2104-2112.
[15]
LIUY F, ZHANGX Y, ZHANGS Y, et al. Part-aware prototype network for few-shot semantic segmentation[C]//Computer Vision-ECCV 2020.Cham:Springer,2020:142-158.
[16]
ZHOUT F, ZHANGM J, ZHAOF,et al. Regional semantic contrast and aggregation for weakly supervised semantic segmentation[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 18-24,2022,New Orleans,LA,USA. IEEE, 2022: 4289-4299.
[17]
WUY C, WUZ H, WUQ Y,et al .Exploring smoothness and class-separation for semi-supervised medical image segmentation[C]//Medical Image Computing and Computer Assisted Intervention-MICCAI 2022. Cham: Springer, 2022: 34-43.
[18]
XUZ, WANGY X, LUD H,et al. All-around real label supervision:cyclic prototype consistency learning for semi-supervised medical image segmentation[J]. IEEE Journal of Biomedical and Health Informatics,2022,26(7): 3174-3184.
[19]
ZHANGZ X, RANR, TIANC N,et al. Self-aware and Cross-sample prototypical learning for semi-supervised medical image segmentation[C]//Medical Image Computing and Computer Assisted Intervention-MICCAI 2023. Cham:Springer,2023:192-201.
[20]
TEJANKARA, ABBASI KOOHPAYEGANIS, PILLAIV,et al .ISD:self-supervised learning by iterative similarity distillation[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV), October 10-17,2021,Montreal,QC,Canada. IEEE,2022: 589-9598.
[21]
YOUC Y, DAIW C, MINY F,et al. Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive distillation[C]//Information Processing in Medical Imaging.Cham:Springer,2023:641-653.
[22]
ZHANGX L, WEIY C, YANGY,et al. SG-one:similarity guidance network for one-shot semantic segmentation[J]. IEEE Transactions on Cybernetics,2020,50(9):3855-3865.
[23]
CLARKK, VENDTB, SMITHK,et al. The cancer imaging archive (TCIA):maintaining and operating a public information repository[J]. Journal of Digital Imaging, 2013, 26(6): 1045-1057.
[24]
XIONGZ H, XIAQ, HUZ Q,et al. A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging[J]. Medical Image Analysis,2021,67:101832.
[25]
WUY C, XUM F, GEZ Y,et al .Semi-supervised left atrium segmentation with mutual consistency training[C]//Medical Image Computing and Computer Assisted Intervention-MICCAI 2021.Cham:Springer,2021:297-306.
[26]
SUJ W, LUOZ M, LIANS,et al .Mutual learning with reliable pseudo label for semi-supervised medical image segmentation[J].Medical Image Analysis,2024,94:103111.
[27]
MAIER-HEINL, EISENMANNM, REINKEA,et al .Why rankings of biomedical image analysis competitions should be interpreted with care[J].Nature Communications,2018,9:5217.