低劑量電腦斷層篩檢降低了 20% 的肺癌死亡率，而廣泛的實行加重了醫生的 負擔並增加了誤診率。基於 CNN 的計算機輔助檢測取得了巨大成功，但仍有減少假 陽性和從理論到臨床實際應用的距離。在這項研究中，我們納入來自成功大學附設 醫院的第一手 電腦斷層影像，從頭開始建立包括臨床標籤轉換的算法，並以解剖學 上更全面的方式修改圖像預處理步驟，以防止預期的切片截斷和元件遺失。我們提 出了一個三維兩階段多尺度檢測器，由一個基於錨優化的區域提取網絡(RPN)和 一個能夠進行多尺度特徵提取的假陽性減少網路(FPR)組成。 我們提出的 RPN 和 二階段模型在知名的 LUNA16 資料集上獲得了 88.1%和 88.9%的競爭性能指標 (CPM)。在更加艱難的成大醫院資料集(the NCKUH dataset)上，PRN 和二階段 模型的 CPM 分別為 76.1%和 76.9%。相比於原來的前處理方法，本研究提出的方法 獲得的 2.2%的 CPM 顯著(p = 0.006)提升。為了解決肺結節檢測任務中的緊迫和特 殊困難，包括由於電腦斷層圖像的性質而沒有自適應池的多尺度分類和具有訓練推 理差異的不同模型的次優集成，提出的預處理方法、新的無參數注意掩碼、基於掩 碼的歸一化和乙狀前值加權集成在實驗中顯示出效率和巨大潛力。

Low-dose CT screening reduces 20% mortality of lung cancer, while the widespread practice burdens physicians' loads and raises the rate of misdiagnosis. CNN-based computer- aided detections have achieved great successes, yet there's room for false-positive reduction and bridging from theory to clinical practice. This study involves the CT scans from the National Cheng Kung University Hospital (NCKUH), establishing an algorithm including clinical label transformation from scratch, and modifying the image preprocessing steps in a more anatomically comprehensive way to prevent unexpected slice truncation and component depletion. We propose a 3D two-stage multi-scale detector consisting of an anchor-optimized Region Proposal Network (RPN) and a False Positive Reduction (FPR) competent for multi-scale feature extraction. The proposed RPN and the entire two-stage model achieved CPMs in the LUNA16 dataset of 88.1% and 88.9%, respectively. In the way more challenging NCKUH in-house dataset, the CPMs are 76.1% in the proposed RPN and 76.9% in the two-stage model, with a significant CPM improvement of 2.2% (p = 0.006) compared to the original preprocessing method. To address urging and specialized difficulties in pulmonary nodule detection tasks, including multi-scale classification without adaptive pooling due to the nature of CT images and suboptimal ensembling of diverse models with training-inference disparities, the proposed methods of preprocessing, the novel parameter-free attention mask, and the pre-sigmoid weighting ensemble show efficiency and great potential in the experiments.

[1] F. Bray, I. Soerjomataram, R. L. Siegel, L. A. Torre and A. Jemal, "Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.," A Cancer J Clin., 2018.
[2] J. Ding, A. Li, Z. Hu and L. Wang, "Accurate pulmonary nodule detection in computed tomography images using deep convolutional neural networks.," in MICCAI, 2017.
[3] Q. Dou, H. Chen, Y. Jin, H. Lin, J. Qin and P.-A. Heng, "Automated pulmonary nodule detection via 3D ConvNets with online sample filtering and Hybrid-Loss residual learning.," in MICCAI, 2017.
[4] F. Liao, M. Liang, X. Hu and S. Song, "Evaluate the malignancy of pulmonary nodules using the 3D deep leaky noisy-OR network.," IEEE Transactions on Neural Networks and Learning Systems, 2019.
[5] H. Peng, H. Sun and Y. Guo, "3D multi-scale deep convolutional neuralnetworks for pulmonary nodule detection.," PLOS ONE, 2020.
[6] A. Setio , "Pulmonary nodule detection in ct images: False positive reduction using multi-view convolutional networks.," IEEE TMI, 2016.
[7] H. Tang, C. Zhang and X. Xie, "NoduleNet: Decoupled false positive reduction for pulmonary nodule detection and segmentation.," in MICCAI, 2019.
[8] W. Zhu, C. Liu, W. Fan and X. Xie, "DeepLung: Deep 3D dual path nets for automated pulmonary nodule detection and classification.," in WACV, 2018.
[9] W.Zhao,.J.Yang,Y.Sun,C.Li,W.Wu,L.Jin,Z.Yang,B.Ni,P.Gao,P.Wang,Y. Hua and M. Li, "3D deep learning from CT scans predicts tumor invasiveness of subcentimeter pulmonary adenocarcinomas.," Cancer Research, 2018.
[10] H. Xie, D. Yang, N. Sun, Z. Chen and Y. Zhang, "Automated pulmonary nodule detection in CT images using deep convolutional neural networks.," Pattern Recognition, 2019.
[11] A. Setio , "Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: The luna16 challenge.," Medical Image Analysis, 2017.
[12] S. Ren, K. He, R. Girshick and J. Sun, "Faster R-CNN: Towards real-time object detection with region proposal networks.," in NIPS, 2015.
[13] O. Ronneberger, P. Fischer and T. Brox, "U-Net: Convolutional networks for biomedical image segmentation.," in MICCAI, 2015.
[14] Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox and O. Ronneberger, "3D U-Net learning dense volumetric segmentation from sparse annotation.," in MICCAI, 2016.
[15] Y. Chen, J. Li, H. Xiao, X. Jin, S. Yan and J. Feng, "Dual path networks.," in NIPS, 2017.
[16] S. B. Shrey, L. Hakim, M. Kavitha, H. W. Kim and T. Kurita, "Transfer learning by cascaded network to identify and classify lung nodules for cancer detection.," Frontiers of Computer Vision, 2020.
[17] J. Long, E. Shelhamer and T. Darrell, "Fully convolutional networks for semantic segmentation.," in CVPR, 2015.
[18] K. He, X. Zhang, S. Ren and J. Sun, "Deep residual learning for image recognition.," in CVPR, 2016.
[19] H. Gao, Z. Liu, L. v. d. Maaten and K. Q. Weinberger, "Densely connected convolutional networks.," in CVPR, 2017.
[20] W. Shen, M. Zhou, F. Yang, C. Yang and J. Tian, "Multi-scale convolutional neural networks for lung nodule classification.," in IPMI, 2015.
[21] S. Harald, "Hinge rank loss and the Area Under the ROC Curve.," in ECML, 2007.
[22] Z. Cai and N. Vasconcelos, "Cascade R-CNN: Delving into high quality object
detection.," in CVPR, 2018.
[23]Y. Li and Y. Fan, "DeepSEED: 3D Squeeze-and-Excitation Encoder-Decoder Convolutional Neural Networks for Pulmonary Nodule Detection," in IEEE International Symposium on Biomedical Imaging, 2020.