肺癌仍是全球癌症相關死亡的主要原因之一，早期偵測對於改善病患預後至關重要。本研究探討YOLO（You Only Look Once）架構與深度學習技術於胸腔電腦斷層影像中肺部結節之檢測與分類應用。利用由多位放射科醫師註解的LIDC-IDRI資料集，我們提出一種改良型的YOLOv11模型，結合多重膨脹注意力機制與重參數化骨幹模組（RepC3、C2PSA），以提升特徵提取與結節定位能力。 
本研究方法著重於將膨脹卷積與注意力模組整合至YOLOv11中，以捕捉多尺度空間特徵，針對小結節與模糊案例等挑戰進行優化。模型透過平均準確率（mAP）、精確度（precision）、召回率（recall）與F1分數等指標進行評估。結果顯示，所提出之YOLO模型在mAP@0.5達到81.34%，F1分數為79.12%，優於YOLOv8與原始YOLOv11模型。此模型在結節分類（良性、不明與惡性）方面亦表現出色，具備降低誤判與漏診風險的潛力。 
本研究強調人工智慧系統在強化肺癌篩檢上的應用潛力，為臨床工作流程提供具擴展性之解決方案。未來方向包括進一步探討三維體積分析、整合臨床資料，以及在資源有限環境下的實地部署，以提升診斷準確性與普及性。 

Lung cancer remains a leading cause of cancer-related deaths globally, with early detection being 
critical for improving patient outcomes. This study explores the application of the YOLO (You Only 
Look Once) architecture and deep learning techniques for the detection and classification of pulmonary 
nodules in chest CT images. Leveraging the LIDC-IDRI dataset, which includes annotated CT scans 
reviewed by multiple radiologists, we propose an enhanced YOLOv11 model incorporating multi
dilation attention mechanisms and reparameterized backbone blocks (RepC3, C2PSA) to improve 
feature extraction and nodule localization. 
Our methodology focuses on augmenting YOLOv11 with dilated convolutions and attention 
modules to capture multi-scale spatial features, addressing challenges such as small nodule detection 
and ambiguous cases. The model was evaluated using metrics including mean Average Precision (mAP), 
precision, recall, and F1-score. Results demonstrate that the proposed YOLO model achieves superior 
performance, with an mAP@0.5 of 81.34% and a balanced F1-score of 79.12%, outperforming baseline 
YOLOv8 and YOLOv11 architectures. The model excels in classifying nodules into benign, ambiguous, 
and malignant categories, offering a robust tool for reducing false positives and missed diagnoses. 
This research highlights the potential of AI-driven systems to enhance lung cancer screening, 
providing a scalable solution for clinical workflows. Future work may explore 3D volumetric analysis, 
integration of clinical metadata, and deployment in resource-limited settings to further advance 
diagnostic accuracy and accessibility. 

Table of Contents
Chapter 1 Introduction	1
1.1 Background and Significance	1
1.2 Introduction of YOLO	2
1.3 Thesis Overview	4
Chapter 2: Related Works	5
2.1 Lung Nodules: Definition, Causes, and Clinical Significance	5
2.2 Computed Tomography (CT) Imaging for Pulmonary Diagnosis	6
2.3 YOLO-Based Detection of Pulmonary Nodules	6
2.4 Lung Nodule Detection and Classification Using YOLO	7
2.5 Lung Nodule Classification Categories	7
Chapter 3: Research Methodology	13
3.1 Overview of YOLOv11 Architecture	13
3.2 Proposed Methodology for Pulmonary Nodule Detection	21
Chapter 4: Experimental Results and Discussions	29
4.1 Dataset and Preprocessing	29
4.2 Training Configuration and Evaluation Metrics	32
4.3 Results and Discussions	35
4.3.1 Original YOLO Experiment Results	35
4.3.2 Ablation Study Results	46
4.3.3 Architectural Comparison: YOLOv11, YOLOv12, and the Proposed Model	49
Chapter 5: Conclusion and Future Works	51
5.1 Conclusion	51
5.2 Future Works	52
References	54

 
List of Figures
Figure 1. Healthy Lung CT Image.	9
Figure 2. Benign Lung Nodule CT Image.	10
Figure 3. Ambiguous Lung Nodule CT Image.	10
Figure 4. Malignant Lung Nodule CT Image.	11
Figure 5. YOLOv11 Backbone Architecture Diagram.	16
Figure 6. YOLOv11 Neck Architecture Diagram.	17
Figure 7. C3K2 Module Diagram.	18
Figure 8. C2PSA Module Diagram.	19
Figure 9. YOLOv11 Architecture.	20
Figure 10. MSDA Module.	25
Figure 11. RepC3 Module. n indicates the number of times it will run RepConv.	26
Figure 12. RepC3 + MSDA Module.	26
Figure 13. The Proposed YOLO Attention Framework Neck.	27
Figure 14. The Proposed YOLO Attention Framework.	28
Figure 15. 	YOLOv11 Metrics and Training/Validation Loss Graphs. YOLOv11 achieves slightly higher early precision, the proposed architecture offers a more balanced and robust performance suitable for clinical deployment.	38
Figure 16. 	Proposed Architecture Metrics and Training/Validation Loss Graphs. The proposed model demonstrates smoother and faster convergence across all loss curves, higher recall, and improved mAP scores—particularly mAP@0.5:0.95—indicating enhanced ability to detect small or ambiguous nodules.	39
Figure 17. YOLOv11 Normalized Confusion Matrix.	40
Figure 18. Proposed Architecture Normalized Confusion Matrix.	41
Figure 19. YOLOv11 Precision-Recall Curve.	42
Figure 20. Proposed Architecture Precision-Recall Curve.	43
Figure 21. 	Validation label visualization for a batch of CT images. Ground truth bounding boxes are annotated with class indices (0: benign, 1: ambiguous, 2: malignant) and corresponding malignancy scores.	44
Figure 22. 	Prediction output of the proposed YOLO model on the same validation batch. The bounding boxes display predicted classes and confidence scores, demonstrating accurate localization and malignancy classification of pulmonary nodules.	45

List of Tables
Table 1. Comparison of Healthy Lung with Different Nodules.	12
Table 2. Dataset Malignancy Interpretation.	31
Table 3. YOLO Training Parameters.	34
Table 4. Hardware Specification.	35
Table 5. 	Ablation study comparing YOLOv8 and YOLOv11 with and without RepC3 and MSDA modules.	48
Table 6. Comparison Results of Different YOLO versions and Proposed version.	50

References 
[1]	World Health Organization, “Cancer,” World Health Organization, Feb. 03, 2025. https://www.who.int/news-room/fact-sheets/detail/cancer
[2]	D. Ardila et al., “End-to-end Lung Cancer Screening with three-dimensional Deep Learning on low-dose Chest Computed Tomography,” Nature Medicine, vol. 25, no. 6, pp. 954–961, May 2019, doi: https://doi.org/10.1038/s41591-019-0447-x.
[3]	F. Islami et al., “Proportion and number of cancer cases and deaths attributable to potentially modifiable risk factors in the United States,” CA: A Cancer Journal for Clinicians, vol. 68, no. 1, pp. 31–54, Nov. 2017, doi: https://doi.org/10.3322/caac.21440.
[4]	H. MacMahon et al., “Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017,” Radiology, vol. 284, no. 1, pp. 228–243, Jul. 2017, doi: https://doi.org/10.1148/radiol.2017161659.
[5]	A. Esteva et al., “A Guide to Deep Learning in Healthcare,” Nature Medicine, vol. 25, no. 1, pp. 24–29, Jan. 2019, doi: https://doi.org/10.1038/s41591-018-0316-z.
[6]	P. J. Mazzone et al., “Components Necessary for High-Quality Lung Cancer Screening,” Chest, vol. 147, no. 2, pp. 295–303, Feb. 2015, doi: https://doi.org/10.1378/chest.14-2500.
[7]	A. A. A. Setio et al., “Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: The LUNA16 challenge,” Medical image analysis, vol. 42, pp. 1–13, 2017, doi: https://doi.org/10.1016/j.media.2017.06.015.
[8]	Q. Dou, H. Chen, L. Yu, J. Qin, and P.-A. Heng, “Multilevel Contextual 3-D CNNs for False Positive Reduction in Pulmonary Nodule Detection,” IEEE Transactions on Biomedical Engineering, vol. 64, no. 7, pp. 1558–1567, Jul. 2017, doi: https://doi.org/10.1109/tbme.2016.2613502.
[9]	W. Shen, M. Zhou, F. Yang, C. Yang, and J. Tian, “Multi-scale Convolutional Neural Networks for Lung Nodule Classification,” Lecture Notes in Computer Science, pp. 588–599, 2015, doi: https://doi.org/10.1007/978-3-319-19992-4_46.
[10]	H. Jiang, H. Ma, W. Qian, and M. Gao, “Detection of Pulmonary Nodules in CT Images Based on YOLOv4,” International Journal of Imaging Systems and Technology, vol. 31, no. 4, pp. 1806–1816, 2021.
[11]	Mehdi Astaraki, G. Yang, Yousuf Zakko, Iuliana Toma-Dasu, Örjan Smedby, and C. Wang, “A Comparative Study of Radiomics and Deep-Learning Based Methods for Pulmonary Nodule Malignancy Prediction in Low Dose CT Images,” Frontiers in Oncology, vol. 11, Dec. 2021, doi: https://doi.org/10.3389/fonc.2021.737368.
[12]	J. Redmon and A. Farhadi, “YOLOv3: An Incremental Improvement,” arxiv.org, Apr. 2018, doi: https://doi.org/10.48550/arXiv.1804.02767.
[13]	None Sakshiwala and M. P. Singh, “A New Framework for multi-scale CNN-based Malignancy Classification of Pulmonary Lung Nodules,” Journal of Ambient Intelligence and Humanized Computing, vol. 14, no. 5, pp. 4675–4683, Sep. 2022, doi: https://doi.org/10.1007/s12652-022-04368-w.
[14]	C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” arXiv (Cornell University), Jul. 2022, doi: https://doi.org/10.48550/arxiv.2207.02696.
[15]	S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137–1149, Jun. 2017, doi: https://doi.org/10.1109/tpami.2016.2577031.
[16]	A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Optimal Speed and Accuracy of Object Detection,” arXiv.org, Apr. 22, 2020. https://www.arxiv.org/abs/2004.10934v1
[17]	J. Mei, M.-M. Cheng, G. Xu, L.-R. Wan, and H. Zhang, “SANet: A Slice-Aware Network for Pulmonary Nodule Detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–1, 2021, doi: https://doi.org/10.1109/tpami.2021.3065086.
[18]	H. Zhang, K. Zu, J. Lu, Y. Zou, and D. Meng, “EPSANet: An Efficient Pyramid Squeeze Attention Block on Convolutional Neural Network,” Lecture notes in computer science, pp. 541–557, Jan. 2023, doi: https://doi.org/10.1007/978-3-031-26313-2_33.
[19]	D. E. Ost and M. K. Gould, “Decision Making in Patients with Pulmonary Nodules,” American Journal of Respiratory and Critical Care Medicine, vol. 185, no. 4, pp. 363–372, Feb. 2012, doi: https://doi.org/10.1164/rccm.201104-0679ci.
[20]	R. M. Tuder and I. Petrache, “Pathogenesis of chronic obstructive pulmonary disease,” Journal of Clinical Investigation, vol. 122, no. 8, pp. 2749–2755, Aug. 2012, doi: https://doi.org/10.1172/jci60324.
[21]	M. Sánchez, M. Benegas, and I. Vollmer, “Management of incidental lung nodules ,” Journal of Thoracic Disease, vol. 10, no. Suppl 22, pp. S2611–S2627, Aug. 2018, doi: https://doi.org/10.21037/jtd.2018.05.86.
[22]	G. D. Rubin, “Computed Tomography: Revolutionizing the Practice of Medicine for 40 Years,” Radiology, vol. 273, no. 2S, pp. S45–S74, Nov. 2014, doi: https://doi.org/10.1148/radiol.14141356.
[23]	N. Horeweg et al., “Characteristics of lung cancers detected by computer tomography screening in the randomized NELSON trial,” American Journal of Respiratory and Critical Care Medicine, vol. 187, no. 8, pp. 848–854, Apr. 2013, doi: https://doi.org/10.1164/rccm.201209-1651OC.
[24]	D. Wormanns and S. Diederich, “Characterization of small pulmonary nodules by CT,” European Radiology, vol. 14, no. 8, May 2004, doi: https://doi.org/10.1007/s00330-004-2335-z.
[25]	S. G. Armato et al., “The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans,” Medical physics, vol. 38, no. 2, pp. 915–31, 2011, doi: https://doi.org/10.1118/1.3528204.
[26]	W. Liu et al., “SSD: Single Shot MultiBox Detector,” Computer Vision – ECCV 2016, vol. 9905, pp. 21–37, 2016, doi: https://doi.org/10.1007/978-3-319-46448-0_2.
[27]	T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollar, “Focal loss for dense object detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–1, 2018, doi: https://doi.org/10.1109/tpami.2018.2858826.
[28]	S. N. Rao, “YOLOv11 Architecture Explained: Next-Level Object Detection with Enhanced Speed and Accuracy,” Medium, Oct. 22, 2024. https://medium.com/@nikhil-rao-20/yolov11-explained-next-level-object-detection-with-enhanced-speed-and-accuracy-2dbe2d376f71