近年來，肺癌逐漸成為國人的重要病症之一，在病發早期醫生能透過低劑量電腦斷層掃描(LDCT)的影像，來捕捉更小的肺結節，能在早期檢測肺癌方面具有潛力，有助於提高治療的成功率。在肺癌早期診斷中，可以過深度學習的方式分析LDCT影像，協助醫生快速且準確的識別肺結節，深度學習技術能更有效的提高肺結節的檢出率，進一步提升患者的治療效果。
在本論文中所使用到的資料集由成大醫院胸腔外科提供，我們透過改良過的模型3D Transfomer UNet進行segmentation，分割出LDCT中肺結節的區域，在經由刪除肺腔室外的肺結節、太小的肺結節以及組織器官，來減少false positive。
在3D Transfomer UNet的模型架構中，以3D UNet為基礎進行改良，並加入transformer的機制，使4mm以上的肺結節準確度，在平均一個病人出現兩顆false positive的情況下，recall從73.5%提升到77.3%，在平均一個病人有十二顆false positive的情況下，recall為91.1%。
儘管本論文所提出的3D Transformer UNet和false positive reduction的方法，在4mm以上的肺結節能有很好的偵測效果，能協助醫生檢測出病人的肺結節位置，但仍然面臨很多挑戰，例如: 4mm以下的肺結節難以偵測、模型所需的硬體資源大、模型預測後有很多false positive產生等，期望在未來能藉由改變輸入模型的解析度來嘗試解決， 使本論文的研究和發展能更加完善，讓其在臨床實踐中發揮更大的作用。

In recent years, lung cancer has gradually become one of the significant health concerns among the population. In the early stages of the disease, the doctors have found potential in utilizing Low-Dose Computed Tomography (LDCT) images to detect lung cancer early by capturing smaller lung nodules that enhance the success rate of treatment. In the early diagnosis of lung cancer, deep learning techniques have been employed to analyze LDCT images, assisting doctors to quickly and accurately identify lung nodules. The application of deep learning technology can effectively increase the detection rate of lung nodules, further improving the effectiveness of patient treatment.
The dataset used in this paper was provided by the Department of Thoracic Surgery at National Cheng Kung University Hospital. We segmented the lung nodules in the LDCT through the improved model 3D Transformer UNet, and deleted the areas outside the lung region, too small lung nodules and tissues and organs to reduce false positive.
In the architecture of the 3D Transformer UNet model, a modified version of the 3D UNet was incorporated with the addition of transformer mechanisms. This enhancement significantly improved the accuracy of lung nodule detection for nodules of 4mm or larger. Under the scenario of an average of two false positives per patient, the recall rate increased from 73.5% to 77.3%. With an average of twelve false positives per patient, the recall rate reached 91.1%.
In the model structure of 3D Transformer UNet, we improve performance based on 3D UNet by transformer mechanism. In the case of two false positive per patient, the recall of lung nodules over 4mm can be increased from 73.5% to 77.3%. In the case of twelve false positive per patient, the recall rate of lung nodules over 4mm is 91.1%。
Although the 3D Transformer UNet and false positive reduction methods proposed in this paper have a good detection effect on lung nodules above 4mm and can assist doctors in detecting the location of lung nodules in patients, they still face many challenges, such as lung nodules below 4mm are difficult to detect, the hardware resources required by the large model and many false positives are generated after model prediction, etc. It is hoped that we can solve the problem in the future by changing the resolution of the input model, making this paper can be improved to allow it to play a important role in clinical practice.

[1] Ö. Çiçek, A. Abdulkadir, S. Lienkamp, T. Brox and O. Ronneberger. 3D U-Net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016
[2] L.C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam. (2018). Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV) (pp. 801-818).
[3] J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang and Y. Zhou (2021). Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306.
[4] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold and S. Gelly (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.
[5] D. Maji, P. Sigedar and M. Singh (2022). Attention Res-UNet with Guided Decoder for semantic segmentation of brain tumors. Biomedical Signal Processing and Control, 71, 103077.
[6] B.C. Kim, J.S. Yoon, J.S. Choi, & H.I. Suk (2019). Multi-scale gradual integration CNN for false positive reduction in pulmonary nodule detection. Neural Networks, 115, 1-10.
[7] J. Mei, M.M. Cheng, G. Xu, L.R. Wan and H. Zhang (2021). SANet: A slice-aware network for pulmonary nodule detection. IEEE transactions on pattern analysis and machine intelligence, 44(8), 4374-4387.
[8] O. Petit, N. Thome, C. Rambour, L. Themyr, T. Collins and L. Soler (2021). U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021
[9] A. Saha, M. Hosseinzadeh and H. Huisman (2021). End-to-end prostate cancer detection in bpMRI via 3D CNNs: effects of attention mechanisms, clinical priori and decoupled false positive reduction. Medical image analysis, 73, 102155.
[10] O. Oktay, J. Schlemper, L.L. Folgoc, M. Lee, M. Heinrich, K. Misawa and D. Rueckert (2018). Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999.
[11] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez and I. Polosukhin (2017). Attention is all you need. Advances in neural information processing systems, 30.
[12] T.Y. Lin, P. Goyal, R. Girshick, K. He and P. Dollár (2017). Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (pp. 2980-2988).
[13] H. Cao, H. Liu, E. Song, G. Ma, X. Xu, R. Jin and C.C. Hung. (2020). A two-stage convolutional neural networks for lung nodule detection. IEEE journal of biomedical and health informatics, 24(7), 2006-2015.
[14] O. Ronneberger, P. Fischer, and T. Brox (2015). U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18 (pp. 234-241). Springer International Publishing.
[15] Z.W. Zhou, M.M.R. Siddiquee, N. Tajbakhsh and J.M. Liang, “UNet++: A Nested U-Net Architecture for Medical Image Segmentation”, In DLMIA, 2018.
[16] F. Milletari, N. Navab and S.A. Ahmadi, “V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation”, In 3DV, 2016.
[17] S. Jadon (2020, October). A survey of loss functions for semantic segmentation. In 2020 IEEE conference on computational intelligence in bioinformatics and computational biology (CIBCB) (pp. 1-7). IEEE.
[18] H. Tang, C. Zhang and X. Xie (2019). Nodulenet: Decoupled false positive reduction for pulmonary nodule detection and segmentation. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part VI 22 (pp. 266-274). Springer International Publishing.
[19] H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y. Iwamoto and J. Wu (2020, May). Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) (pp. 1055-1059). IEEE.
[20] G. Xu, X. Wu, X. Zhang, and X. He (2021). Levit-unet: Make faster encoders with transformer for medical image segmentation. arXiv preprint arXiv:2107.08623.
[21] D. Bolya, C. Zhou, Fanyi Xiao and Yong Jae Lee, “YOLACT: Real-time Instance Segmentation”, In ICCV, 2019.
[22] K. He, G. Gkioxari, P. Dollár and R. Girshick (2017). Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 2961-2969).
[23] “Lung cancer”, WHO世界衛生組織, 12 06 2023. Available:https://www.who.int/news-room/fact-sheets/detail/lung-cancer
[24] “衛生福利部國民健康署肺癌早期偵測計畫”, 衛生福利部, 02 24 2023. Available:https://www.laws.taipei.gov.tw/lawatt/Law/A040170071010500-20230726-3000-065.pdf