肺癌已連續多年蟬聯臺灣的死因之首，因此，衛生福利部開始為高風險族群提供每兩年一次的低劑量電腦斷層掃描（LDCT）肺癌篩檢，因為此政策的關係，讓醫生多了很多CT片要看，對精準的肺部電腦輔助診斷（CAD）系統的需求也隨之增加。在典型的肺結節偵測電腦輔助診斷系統中，會分成兩個階段：1) 偵測候選肺結節與 2) 減少偽陽性。本研究專注在第二階段的減少偽陽性，目標是基於候選肺結節的影像，在保持第一階段偵測到的真實結節的同時，盡可能地減少偽陽性。關於減少肺結節的偽陽性的研究，很多都是基於卷積神經網路（CNN）來做的，不過最近Transformer被廣泛應用在許多電腦視覺的任務中，並且得到不錯的結果，因此，本研究與國立成功大學醫學院附設醫院（簡稱成大醫院）的胸腔外科合作，提出了一個改良的3D CNN-Transformer模型。該模型結合了CNN與Transformer的優點，且因為肺結節本質上是3D的，所以在CNN的部分使用了3D卷積模組萃取3D的局部特徵，Transformer的部分則負責捕捉3D局部特徵之間的關係。本研究提出的模型會對3D候選結節進行詳細分析，它萃取候選肺結節在多重尺度的特徵，將它們轉換為特徵描述符（feature descriptor），並分析這些描述符，以準確的判斷候選肺結節屬於陽性（真實的結節），還是屬於偽陽性。根據國立成功大學機器人實驗室開發的肺結節偵測模型所產生的資料進行測試，此模型對直徑大於4毫米的結節的召回率（Recall）為 87.6%，對所有大小的結節的召回率為 82.5%。當與偵測模型整合時，考慮到被偵測模型忽略掉的結節，這個系統對成大醫院的資料集裡直徑大於4毫米的結節可以達到82.5%的召回率，對所有大小的結節的召回率則為66.7%。本研究所提出的模型亦整合在機器人實驗室所開發的聯邦學習系統裡面，未來，它將使用來自更多醫院的醫學影像資料進行訓練，產生更強大的模型，能夠更精準的減少肺結節偵測的偽陽性。

Lung cancer has been the leading cause of cancer death for many years in Taiwan. Thus, the Taiwan Ministry of Health and Welfare (MOHW) has started providing biannual low-dose computed tomography (LDCT) lung screening for high-risk groups. It would give doctors many chest CT scans to review and creates the need for accurate computer-aided diagnosis (CAD) for the lungs. A typical CAD system for lung nodule detection has two stages: 1) nodule candidate detection and 2) false positive reduction. This work focuses on the second stage. The objective is to reduce as many false positives as possible while maintaining the true nodules found in the first stage, based on the images of nodule candidates. Many studies on false positive reduction have been done based on CNN. Recently, Transformer was widely used in many computer vision tasks, showing promising results. Thus, this work proposed a Modified 3D CNN-Transformer Model in collaboration with the Division of Thoracic Surgery, National Cheng Kung University Hospital (NCKUH). This model combines the advantages of CNN and Transformer. Since lung nodules are 3D in nature, CNN uses 3D convolution blocks to extract 3D local features. Transformer captures relationships of 3D local features. The proposed model performs detailed analyses of 3D nodule candidates. It extracts multi-scale features of the candidate, encodes them as feature descriptors, and analyzes the features in the descriptors to accurately determine if the candidate is a True Positive or False Positive. The proposed model achieved a recall rate of 87.6% for the nodules with diameters larger than 4 mm, and 82.5% for all the nodules, based on the data from the nodule candidate detection model made by Robotics Lab., National Cheng Kung University (NCKU). When integrating with the nodule candidate detection model made, considering the missing in it, the system can achieve a recall rate of 82.5% for the nodules with diameters larger than 4 mm and 66.7% for all the nodules in the data from NCKUH. The proposed model is also integrated with the federated learning system made by Robotics Lab. It will be trained with more medical images from more hospitals in the future, delivering a more robust false positive reduction model.

[1] "111年國人死因統計結果-衛生福利部," 衛生福利部, 12 06 2023. [Online]. Available: https://www.mohw.gov.tw/cp-16-74869-1.html. [Accessed 31 07 2023].
[2] "Lung cancer," WHO, 26 06 2023. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/lung-cancer. [Accessed 31 07 2023].
[3] P. J. Mazzone and L. Lam, "Evaluating the Patient With a Pulmonary Nodule: A Review," JAMA, vol. 327, no. 3, pp. 264-273, 2022.
[4] T. Hunger, E. Wanka-Pail, G. Brix and J. Griebel, "Lung Cancer Screening with Low-Dose CT in Smokers: A Systematic Review and Meta-Analysis," Diagnostics, vol. 11, no. 6, 2021.
[5] "衛生福利部國民健康署-肺癌早期偵測計畫," 衛生福利部國民健康署, 31 05 2022. [Online]. Available: https://www.hpa.gov.tw/Pages/List.aspx?nodeid=4619. [Accessed 01 08 2023].
[6] P. F. Pinsky, "Assessing The Benefits and Harms of Low-Dose Computed Tomography Screening for Lung Cancer," Lung Cancer Management, vol. 3, no. 6, pp. 491-498, 2015.
[7] K. Greenway, A. Murphy and G. Frank, "Hounsfield unit. Reference Article, Radiopaedia.org," 23 03 2015. [Online]. Available: https://radiopaedia.org/articles/38181. [Accessed 11 08 2023].
[8] "DICOM Standartd Browser: All CIODs," Innolitics, 2016. [Online]. Available: https://dicom.innolitics.com/ciods. [Accessed 06 08 2023].
[9] D. Richfield, "Medical gallery of David Richfield 2014," WikiJournal of Medicine, vol. 1, no. 2, 2014.
[10] B.-C. Kim, J. S. Yoon, J.-S. Choi and H.-I. Suk, "Multi-Scale Gradual Integration CNN for False Positive Reduction in Pulmonary Nodule Detection," Neural Networks, vol. 115, pp. 1-10, 2019.
[11] "Lung Rads | American College of Radiology," 2019. [Online]. Available: https://www.acr.org/Clinical-Resources/Reporting-and-Data-Systems/Lung-Rads. [Accessed 02 08 2023].
[12] M. J. Sheller, B. Edwards, G. A. Reina, J. Martin, S. Pati, A. Kotrotsou, M. Milchenko, W. Xu, D. Marcus, R. R. Colen and S. Bakas, "Federated Learning in Medicine: Facilitating Multi-Institutional Collaborations without Sharing Patient Data," Scientific Reports, vol. 10, 2020.
[13] M. P. Saradhi and T. V, "Multiscale CNN with Compound Fusions for False Positive Reduction in Lung Nodule Detection," Artificial Intelligence in Medicine, vol. 113, 2021.
[14] C. Niu and G. Wang, "Unsupervised Contrastive Learning Based Transformer for Lung Nodule Detection," Physics in Medicine & Biology, vol. 67, no. 20, 2022.
[15] Y. Dai, Y. Gao and F. Liu, "TransMed: Transformers Advance Multi-Modal Medical," Diagnostics, vol. 11, no. 8, 2021.
[16] J. Jang and D. Hwang, "M3T: Three-Dimensional Medical Image Classifier Using Multi-Plane and Multi-slice Transformer," in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
[17] C.-F. Chen, Q. Fan and R. Panda, "CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification," in 2022 IEEE/CVF International Conference on Computer Vision (ICCV), 2021.
[18] S. Woo, J. Park, J.-Y. Lee and I. S. Kweon, "CBAM: Convolutional Block Attention Module," in Computer Vision – ECCV 2018: 15th European Conference, 2018.
[19] T.-Y. Lin, P. Goyal, R. Girshick, K. He and P. Dollar, "Focal Loss for Dense Object Detection," in 2022 IEEE/CVF International Conference on Computer Vision (ICCV), 2017.
[20] S. Ioffe and C. Szegedy, "Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift," in 32nd International Conference on Machine Learning (PMLR), 2015.
[21] X. Glorot, A. Bordes and Y. Bengio, "Deep Sparse Rectifier Neural Networks," in Fourteenth International Conference on Artificial Intelligence and Statistics, 2011.
[22] Z. Yao, Y. Cao, Y. Lin, Z. Liu, Z. Zhang and H. Hu, "Leveraging Batch Normalization for Vision Transformers," in 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021.
[23] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit and N. Houlsby, "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale," in 9th International Conference on Learning Representations, ICLR 2021, 2021.
[24] J. L. Ba, J. R. Kiros and G. E. Hinton, "Layer Normalization," arXiv:1607.06450 [stat.ML], 2016.
[25] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser and I. Polosukhin, "Attention is All you Need," Advances in Neural Information Processing Systems 30 (NIPS 2017), vol. 30, 2017.
[26] D. Hendrycks and K. Gimpel, "Gaussian Error Linear Units (GELUs)," arXiv:1606.08415 [cs.LG], 2016.
[27] G. Polat, U. Halici and Y. S. Dogrusoz, "False Positive Reduction in Lung Computed Tomography Images using Convolutional Neural Networks," arXiv:1811.01424 [cs.CV], 2018.
[28] Z. Wu, R. Ge, G. Shi, L. Zhang, Y. Chen, L. Luo, Y. Cao and H. Yu, "MD-NDNet: A Multi-Dimensional Convolutional Neural Network for False-Positive Reduction in Pulmonary Nodule Detection," Physics in Medicine & Biology, vol. 65, no. 23, 2020.
[29] Y. R. Chillakuru, K. Kranen, V. Doppalapudi, Z. Xiong, L. Fu, A. Heydari, A. Sheth, Y. Seo, T. Vu and J. H. Sohn, "High Precision Localization of Pulmonary Nodules on Chest CT Utilizing Axial Slice Number Labels," BMC Medical Imaging, vol. 21, 2021.
[30] L. Sun, Z. Wang, H. Pu, G. Yuan, L. Guo, T. Pu and Z. Peng, "Attention-Embedded Complementary-Stream CNN for False Positive Reduction in Pulmonary Nodule Detection," Computers in Biology and Medicine, vol. 113, 2021.
[31] Z. Sun, S. Cao, Y. Yang and K. Kitani, "Rethinking Transformer-based Set Prediction for Object Detection," in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.
[32] J. Yang, H. Deng, X. Huang, B. Ni and Y. Xu, "Relational Learning Between Multiple Pulmonary Nodules via Deep Set Attention Transformers," in 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), 2020.
[33] Ö. Çiçek, A. Abdulkadir, S. 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, 2016.