作者投稿
专家审稿
编辑办公
邮件订阅
RSS
Semi-supervised Histopathological Image Segmentation Method Based on Multi-task Learning
-
摘要: 病理图像自动分割是计算机辅助诊断技术的重要组成部分，可降低病理科医师工作负担，提高诊断效率和准确性。本文介绍一种结合多任务学习的半监督病理图像分割方法。该方法基于半监督的方式同时进行癌症区域图像分割与分类，即首先基于极少量像素级标注图像对分割网络进行训练，然后结合图像级标注图像同时完成图像分割及分类。在网络训练过程中，通过此2个任务的交替迭代以优化网络参数，降低了深度学习模型对图像标注的依赖性。在此基础上，模型引入了动态加权交叉熵损失函数，可利用分类预测概率值自动完成每个像素的权重分配，以提高分割网络对预测概率值较低目标区域的关注度。该策略可有效保持癌症区域的细节信息，经验证可在像素标注数据量不足的情况下对乳腺癌病理图像获得良好的癌症区域分割结果。Abstract: Automatic segmentation of histopathological image is an important step of computer-aided diagnosis, which can reduce the workload of pathologists and improve the efficiency and diagnosis accuracy. This paper introduces a semi-supervised histopathological image segmentation method combined with multi-task learning. This method could simultaneously segment cancer region and classify image using semi-supervised method. Firstly, we used a limited number of pixel-level labels to train a segmentation network, and then the segmentation network and a classification network were trained using some image-level labels. The network parameters were optimized by alternating iterative method in the process of the network training. This method could reduce the annotation cost compared with the standard supervised method for deep learning model. Furthermore, we introduced a dynamically weighted cross entropy loss function to train the network, which could automatically allocate the weight of each pixel by using the probability of classification prediction. This strategy could promote the segmentation network to pay attention to some target regions with the low probability of classification prediction. Therefore, the details of the cancer regions could be preserved. Experimental results on the breast cancer histopathological image verified that our method outperformed other state-of-the-arts on the cancer segmentation performance under the condition of insufficient pixel-level label data.作者贡献：曾黎负责试验方案设计及论文撰写；汤红忠指导论文撰写修订；王蔚、谢明健负责模型的代码调试及参数优化；吴勇军负责数据整理。利益冲突：所有作者均声明不存在利益冲突
-
图 1 不同级别标注的病理图像
A.切片级标注，蓝色方框为正常组织区域，红色方框为癌症组织区域；B.切片级标注中正常组织区域与癌症区域对应的图像级标注；C.图像级标注中癌症区域对应的像素级标注
表 1 β值对网络分割性能的影响
β DSC IOU 1 0.8431 0.7288 2 0.8437 0.7296 3 0.8479 0.7360 4 0.8450 0.7316 5 0.8423 0.7276 6 0.8419 0.7269 IOU：交并比；DSC：Dice相似系数 
下载: 导出CSV
表 2 α值对网络分割性能的影响
α DSC IOU 0.1 0.8445 0.7309 0.2 0.8473 0.7351 0.5 0.8469 0.7344 0.8 0.8471 0.7347 1.0 0.8479 0.7360 IOU、DSC：同表 1
下载: 导出CSV
表 3 不同损失函数对分割性能的影响
损失函数 DSC IOU CE+CE 0.8411 0.7258 CE+FW 0.8415 0.7263 FW+FW 0.8425 0.7279 FW+DWF 0.8438 0.7298 DWF+DWF 0.8479 0.7360 IOU、DSC：同表 1；CE：无加权的交叉熵损失函数；FW：固定权重的交叉熵损失函数；DWF：动态加权的交叉熵损失函数
下载: 导出CSV
表 4 不同像素级标注量对分割任务DSC的影响
组别 0 1 5 10 20 25 30 40 50 75 100 S组 0.596 0.641 0.648 0.728 0.803 0.836 0.841 0.849 0.853 0.859 0.868 S+C组 0.685 0.722 0.731 0.831 0.835 0.839 0.844 0.854 0.855 0.860 0.871 DSC：同表 1
下载: 导出CSV
表 5 不同像素级标注量对分割任务IOU的影响
组别 0 1 5 10 20 25 30 40 50 75 100 S组 0.425 0.472 0.480 0.573 0.671 0.719 0.727 0.739 0.745 0.753 0.767 S+C组 0.522 0.566 0.578 0.692 0.718 0.724 0.731 0.746 0.748 0.755 0.773 IOU：同表 1
下载: 导出CSV
-
[1] Miller KD, Nogueira L, Mariotto AB, et al. Cancer treatment and survivorship statistics, 2019[J]. CA Cancer J Clin, 2019, 69: 363-385. doi: 10.3322/caac.21565 [2] Huang PW, Lee CH. Automatic classification for patholo-gical prostate images based on fractal analysis[J]. IEEE Trans Med Imaging, 2009, 2: 1037-1050. [3] Feng R, Liu X, Chen J, et al. A deep learning approach for colonoscopy pathology WSI analysis: accurate segmentation and classification[J]. IEEE J Biomed Health Inform, 2020, 25: 3700-3708. [4] Xu J, Cai C, Zhou Y, et al. Multi-tissue partitioning for whole slide images of colorectal cancer histopathology images with deeptissue net[C]. European Congress on Digital Pathology. Springer, 2019: 100-108. [5] Mehta S, Mercan E, Bartlett J, et al. Y-Net: joint segmentation and classification for diagnosis of breast biopsy images[C]. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2018: 893-901. [6] Xu J, Zhou C, Lang B, et al. Deep learning for histopathological image analysis: towards computerized diagnosis on cancers[C]. Deep Learning and Convolutional Neural Networks for Medical Image Computing, 2017: 73-95. [7] Iizuka O, Kanavati F, Kato K, et al. Deep learning models for histopathological classification of gastric and colonic epithelial tumours[J]. Sci Rep, 2020, 10: 1504. doi: 10.1038/s41598-020-58467-9 [8] Ye J, Luo Y, Zhu C, et al. Breast cancer image classification on WSI with spatial correlations[C]. ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019: 1219-1223. [9] Xue Y, Ray N. Cell detection in microscopy images with deep convolutional neural network and compressed sensing[J]. arXiv, 2017. https://arxiv.org/abs/1708.03307v3 .[10] Saleh F, Aliakbarian MS, Salzmann M, et al. Built-in foreground/background prior for weakly-supervised semantic segmentation[C]. European Conference on Computer Vision, 2016: 413-432. [11] Lempitsky V, Kohli P, Rother C, et al. Image segmentation with a bounding box prior[C]. 2009 IEEE 12th International Conference on Computer Vision, 2009: 277-284. [12] Zhou B, Khosla A, Lapedriza A, et al. Learning deep features for discriminative localization[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 2921-2929. [13] Jia Z, Huang X, Eric I, et al. Constrained deep weak supervision for histopathology image segmentation[J]. IEEE Trans Med Imaging, 2017, 36: 2376-2388. doi: 10.1109/TMI.2017.2724070 [14] Chen Z, Chen Z, Liu J, et al. Weakly supervised histopathology image segmentation with sparse point annotations[J]. IEEE J Biomed Health Inform, 2020, 25: 1673-1685. [15] Gildenblat J, Klaiman E. Self-supervised similarity learning for digital pathology[J]. arXiv, 2019. https://arxiv.org/abs/1905.08139v3 .[16] Srinidhi CL, Kim SW, Chen FD, et al. Self-supervised driven consistency training for annotation efficient histopathology image analysis[J]. Med Image Anal, 2022, 75: 102256. doi: 10.1016/j.media.2021.102256 [17] Xu G, Song Z, Sun Z, et al. Camel: A weakly supervised learning framework for histopathology image segmentation[C]. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019: 10682-10691. [18] Lerousseau M, Vakalopoulou M, Classe M, et al. Weakly supervised multiple instance learning histopathological tumor segmentation[C]. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2020: 470-479. [19] Yang Y, Yang Y, Yuan Y, et al. Detecting helicobacter pylori in whole slide images via weakly supervised multi-task learning[J]. Multimed Tools Appl, 2020, 79: 26787-26815. doi: 10.1007/s11042-020-09185-x [20] Ciga O, Martel AL. Learning to segment images with classification labels[J]. Med Image Anal, 2021, 68: 101912. doi: 10.1016/j.media.2020.101912 [21] Mehta S, Rastegari M, Caspi A, et al. Espnet: Efficient spatial pyramid of dilated convolutions for semantic segmentation[C]. Proceedings of the European Conference on Computer Vision (ECCV), 2018: 552-568. [22] Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift[C]. International Conference on Machine Learning, 2015: 448-456. [23] Glorot X, Bordes A, Bengio Y. Deep sparse rectifier neural networks[C]. Proceedings of the 14th International Conference on Artificial Intelligence and Statistics, 2011: 315-323. [24] Msonda P, Uymaz SA, Karaağaç SS. Spatial pyramid pooling in deep convolutional networks for automatic tuberculosis diagnosis[J]. Trait Signal, 2020, 37: 1075-1084. doi: 10.18280/ts.370620 [25] 邓仕俊, 汤红忠, 曾黎, 等. 基于多尺度特征感知的胸腔图像危及器官分割[J]. 中国生物医学工程学报, 2021, 40: 701-711. https://www.cnki.com.cn/Article/CJFDTOTAL-ZSWY202106007.htm [26] Abraham N, Khan NM. A novel focal tversky loss function with improved attention u-net for lesion segmentation[C]. 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), 2019: 683-687. [27] Fidon L, Li W, Garcia-Peraza-Herrera LC, et al. General-ised wasserstein dice score for imbalanced multi-class segmentation using holistic convolutional networks[C]. International Conference on Medical Image Computing and Computer-assisted Intervention, 2017: 64-76. [28] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015: 234-241. [29] Zhang Z, Sabuncu M. Generalized cross entropy loss for training deep neural networks with noisy labels[J]. Adv Neural Inf Process Syst, 2018, 31: 8792-8802. [30] Celik Y, Talo M, Yildirim O, et al. Automated invasive ductal carcinoma detection based using deep transfer learning with whole-slide images[J]. Pattern Recogn Lett, 2020, 133: 232-239. doi: 10.1016/j.patrec.2020.03.011 -
点击查看大图
图(8) / 表(7)
计量
- 文章访问数: 427
- HTML全文浏览量: 80
- PDF下载量: 46
- 被引次数: 0
