由於傳統的腫瘤切割方式是相當費時費力的，發展一個有效的自動腫瘤分割模型是相當重要的。然而，過往大多數的模型都是採用全監督式的方法來訓練的，全監督式的方法需要大量的像素等級標註資料，而像素等級的標註資料卻是相當昂貴的。因此，我們提出一個只需要圖像等級的標註資料和少量的像素等級標註資料的方法來訓練我們的肝臟腫瘤切割模型。我們提出的方法包含兩個元件：對比損失函數和偽標籤產生策略。我們的對比損失函數是採用多重正樣本損失函數，多重正樣本損失函數能夠拉近相同圖像標註的投影特徵距離，並同時推開不同圖像標註的投影特徵距離。我們的偽標籤產生策略採用師徒架構，將老師網路的高信心預測結果做為偽標籤來訓練學生網路。老師網路會在學生網路表現的比老師網路還要好的時候更新，來確保偽標籤的品質。我們在Liver Tumor Segmentation (LiTS)資料集與Kidney Tumor Segmentation (KiTS)資料集中展現我們的實驗成果。我們提出的方法超越了以往的方法，並在LiTS資料集達到了58.93\%的dice score，且距離使用大量像素等級標註資料的全監督式訓練方式僅有1.38\%的距離。

Due to traditional tumor segmentation's time and expertise cost, it is essential to develop an effective automated liver tumor segmentation model. Most works approach the problem in a fully supervised manner, requiring numerous pixel-level annotated data, which is expensive to procure. Our work uses only a small amount of pixel-level annotated data by supplementing it with image-level annotated data to train our liver tumor segmentation model in a weakly and semi-supervised manner. We propose two components for our tumor segmentation approach: a contrastive loss and a pseudo-labeling strategy. Our contrastive loss is the Multiple Positive Instance Contrastive (MPIC) loss which pulls the projected features of the same image labels together while simultaneously pushing the projected features of different image labels away. Our pseudo-labeling strategy uses a teacher-student architecture wherein the high-confidence prediction of the teacher network is used as the pseudo-labels for the student network. The teacher network is updated whenever the student outperforms it to ensure quality pseudo-labels. We demonstrate our model's capabilities with experiments on the Liver Tumor Segmentation (LiTS) dataset and Kidney Tumor Segmentation (KiTS) dataset. Our proposed method outperforms previous works and achieves a dice score of 58.93\% on LiTS dataset, with only a 1.38\% gap compared to the fully supervised approach where pixel-level annotation of all the data is available.

[1] T. A. C. S. medical and editorial content team, “Key statistics about liver cancer,” 2021.
[2] Y. Tang, Y. Tang, Y. Zhu, J. Xiao, and R. M. Summers, “E2net: An edge enhanced network for accurate liver and tumor segmentation on CT scans,” in Medical Image Computing and Computer Assisted Intervention ­ MICCAI 2020 ­ 23rd International Conference, Lima, Peru, October 4­8, 2020, Proceedings, Part IV (A. L. Martel, P. Abolmaesumi, D. Stoyanov, D. Mateus, M. A. Zuluaga, S. K. Zhou, D. Racoceanu, and L. Joskowicz, eds.), vol. 12264 of Lecture Notes in Computer Science, pp. 512–522, Springer, 2020.
[3] J. M. J. Valanarasu, V. A. Sindagi, I. Hacihaliloglu, and V. M. Patel, “Kiu­net: Overcomplete convolutional architectures for biomedical image and volumetric segmentation,” arXiv preprint arXiv:2010.01663, 2020.
[4] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discrimina­ tive localization,” in Computer Vision and Pattern Recognition, 2016.
[5] Y. Wei, H. Xiao, H. Shi, Z. Jie, J. Feng, and T. S. Huang, “Revisiting dilated convolution: A sim­ ple approach for weakly­ and semi­ supervised semantic segmentation,” CoRR, vol. abs/1805.04574, 2018.
[6] J. Lee, E. Kim, S. Lee, J. Lee, and S. Yoon, “Ficklenet: Weakly and semi­supervised semantic image segmentation\\using stochastic inference,” CoRR, vol. abs/1902.10421, 2019.
[7] T. Chen, S. Kornblith, M. Norouzi, and G. E. Hinton, “A simple framework for contrastive learning of visual representations,” CoRR, vol. abs/2002.05709, 2020.
[8] R. Hadsell, S. Chopra, and Y. LeCun, “Dimensionality reduction by learning an invariant mapping,” in 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), vol. 2, pp. 1735–1742, 2006.
[9] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual repre­ sentation learning,” arXiv preprint arXiv:1911.05722, 2019.
[10] Z. Chen, R. Zhang, G. Zhang, Z. Ma, and T. Lei, “Digging into pseudo label: A low­budget approach for semi­supervised semantic segmentation,” IEEE Access, vol. 8, pp. 41830–41837, 2020.
[11] P. Bilic, P. F. Christ, E. Vorontsov, G. Chlebus, H. Chen, Q. Dou, C. Fu, X. Han, P. Heng, J. Hesser, S. Kadoury, T. K. Konopczynski, M. Le, C. Li, X. Li, J. Lipková, J. S. Lowengrub, H. Meine, J. H. Moltz, C. Pal, M. Piraud, X. Qi, J. Qi, M. Rempfler, K. Roth, A. Schenk, A. Sekuboyina, P. Zhou, C. Hülsemeyer, M. Beetz, F. Ettlinger, F. Grün, G. Kaissis, F. Lohöfer, R. Braren, J. Holch, F. Hof­ mann, W. H. Sommer, V. Heinemann, C. Jacobs, G. E. H. Mamani, B. van Ginneken, G. Chartrand, A. Tang, M. Drozdzal, A. Ben­Cohen, E. Klang, M. M. Amitai, E. Konen, H. Greenspan, J. Moreau,A. Hostettler, L. Soler, R. Vivanti, A. Szeskin, N. Lev­Cohain, J. Sosna, L. Joskowicz, and B. H. Menze, “The liver tumor segmentation benchmark (lits),” CoRR, vol. abs/1901.04056, 2019.
[12] N. Heller, N. Sathianathen, A. Kalapara, E. Walczak, K. Moore, H. Kaluzniak, J. Rosenberg, P. Blake, Z. Rengel, M. Oestreich, J. Dean, M. Tradewell, A. Shah, R. Tejpaul, Z. Edgerton, M. Peterson, S. Raza, S. Regmi, N. Papanikolopoulos, and C. Weight, “The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes,” 2020.
[13] X. Wang, S. Han, Y. Chen, D. Gao, and N. Vasconcelos, “Volumetric attention for 3d medical im­ age segmentation and detection,” Medical Image Computing and Computer Assisted Intervention – MICCAI 2019, p. 175–184, 2019.
[14] Y.­C. Liu, D. S. Tan, J.­C. Chen, W.­H. Cheng, and K.­L. Hua, “Segmenting hepatic lesions using residual attention u­net with an adaptive weighted dice loss,” in 2019 IEEE International Conference on Image Processing (ICIP), pp. 3322–3326, 2019.
[15] Y.­C. Liu, M. Shahid, W. Sarapugdi, Y.­X. Lin, J.­C. Chen, and K.­L. Hua, “Cascaded atrous dual attention u­net for tumor segmentation,” Multimedia Tools and Applications, Oct 2020.
[16] C.­C. Hsu, K.­J. Hsu, C.­C. Tsai, Y.­Y. Lin, and Y.­Y. Chuang, “Weakly supervised instance segmen­ tation using the bounding box tightness prior,” in Advances in Neural Information Processing Systems (H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché­Buc, E. Fox, and R. Garnett, eds.), vol. 32, Curran Associates, Inc., 2019.
[17] H. Kervadec, J. Dolz, S. Wang, E. Granger, and I. B. Ayed, “Bounding boxes for weakly supervised segmentation: Global constraints get close to full supervision,” CoRR, vol. abs/2004.06816, 2020.
[18] J. Zhang, X. Yu, A. Li, P. Song, B. Liu, and Y. Dai, “Weakly­supervised salient object detection via scribble annotations,” CoRR, vol. abs/2003.07685, 2020.
[19] Y. Wei and S. Ji, “Scribble­based weakly supervised deep learning for road surface extraction from remote sensing images,” CoRR, vol. abs/2010.13106, 2020.
[20] X. Chen, H. Fan, R. Girshick, and K. He, “Improved baselines with momentum contrastive learning,” arXiv preprint arXiv:2003.04297, 2020.
[21] K. Sohn, D. Berthelot, C.­L. Li, Z. Zhang, N. Carlini, E. D. Cubuk, A. Kurakin, H. Zhang, and C. Raffel, “Fixmatch: Simplifying semi­supervised learning with consistency and confidence,” arXiv preprint arXiv:2001.07685, 2020.
[22] Y. Ouali, C. Hudelot, and M. Tami, “Semi­supervised semantic segmentation with cross­consistency training,” in The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.
[23] L. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder­decoder with atrous separable convolution for semantic image segmentation,” CoRR, vol. abs/1802.02611, 2018.
[24] H. Zhang, C. Wu, Z. Zhang, Y. Zhu, Z. Zhang, H. Lin, Y. Sun, T. He, J. Muller, R. Manmatha, M. Li, and A. Smola, “Resnest: Split­attention networks,” arXiv preprint arXiv:2004.08955, 2020.
[25] A. van den Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” CoRR, vol. abs/1807.03748, 2018.
[26] D. Qin, J.­J. Bu, Z. Liu, X. Shen, S. Zhou, J.­J. Gu, Z.­H. Wang, L. Wu, and H.­F. Dai, “Efficient med­ ical image segmentation based on knowledge distillation,” IEEE Transactions on Medical Imaging, pp. 1–1, 2021.
[27] X. Li, H. Chen, X. Qi, Q. Dou, C. Fu, and P. Heng, “H­denseunet: Hybrid densely connected unet for liver and liver tumor segmentation from CT volumes,” CoRR, vol. abs/1709.07330, 2017.