麴黴屬 (Aspergillus) 是真菌界中極具多樣性和重要性的一類，已確認的物種超過900種。一些 Aspergillus 物種對人類和動物具有致病性，導致各種感染性疾病，如常見的 Aspergillus fumigatus 引起的肺部感染。除了致病性外，一些 Aspergillus 物種還在工業和生物學領域有重要應用價值，例如黑麴黴(Aspergillus niger)常用於生產酵素和食品添加劑。

然而，在傳統的真菌鑑定方法耗時且容易受到主觀偏見的影響，所以在醫學實驗室中，對 Aspergillus 的鑑定仍然是一項挑戰。透過先前 YOLO 物件偵測演算法的研究結果於臨床實際應用發現，儘管在驗證集及測試集中表現出色，但實際應用於臨床後效果未達預期水準，所以本研究旨在驗證先前模型訓練的資料型態可能存在的問題，導致後續模型性能產生高估的影響，並提出了拆分工作號和集成策略等方法來提升模型的預測能力以及可評估性。研究結果顯示，先前研究中資料若未拆分工作號進行模型訓練會存在著資料洩的問題導致模型高估。拆分工作號的模型在全新的測試集中表現穩定，但未拆分工作號的方式也能在全新的測試集中取得相當的效果。進一步地，我們通過生成偽標籤和增量學習等方法，成功提升了模型的性能和預測準確度，並大幅減少了人工標記資料的成本。

The genus Aspergillus is a highly diverse and significant group within the fungal kingdom, with over 900 identified species. Some Aspergillus species are pathogenic to humans and animals, causing various infectious diseases, such as the common lung infections caused by Aspergillus fumigatus. In addition to their pathogenicity, some Aspergillus species have important applications in industrial and biological fields. For example, Aspergillus niger is commonly used in the production of enzymes and food additives.

However, traditional fungal identification methods are time-consuming and prone to subjective bias, making the identification of Aspergillus a challenge in medical laboratories. Previous research on the YOLO object detection algorithm for clinical applications revealed that although the model performed well on validation and test sets, its effectiveness in clinical applications did not meet expectations. This study aims to verify potential issues with the data types used in previous model training, which may have led to an overestimation of model performance. The study proposes methods such as splitting work numbers and ensemble strategies to improve the model's predictive capability and evaluability. The results show that the previous study's data, if not split by work numbers during model training, had data leakage issues leading to overestimated performance. Models trained with split work numbers performed stably on new test sets, whereas models without this split also achieved comparable results on new test sets. Furthermore, we successfully enhanced the model's performance and prediction accuracy through methods such as generating pseudo-labels and incremental learning, significantly reducing the cost of manual data labeling.

論文審定書 i
誌謝 ii
摘要 iii
Abstract iv
Chapter 1 緒論 1
1.1 研究背景 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 研究動機與目的 . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 文獻回顧. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 研究流程與架構 . . . . . . . . . . . . . . . . . . . . . . . . . . .2
Chapter 2 資料集與資料前處理 4
2.1 資料集介紹 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 資料前處理 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 資料集類別數量統計與分析. . . . . . . . . . . . . . . . 7
2.2.2 圖像標註 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.3 資料切分及命名(工作號有無區分、後續資料及使用分配) . . . . . 10
Chapter 3 研究方法 15
3.1 環境建置(Ultralytics) . . . . . . . . . . . . . . . . . . . . . . 15
3.2 參數設置 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.1 Mosaic數據增強技術. . . . . . . . . . . . . . . . . . . . . 18
3.3 模型架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3.1 YOLOv8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3.2 YOLOv9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 模型評估機制 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29
3.4.1 損失函數 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29
3.4.2 平均精度均值 . . . . . . . . . . . . . . . . . . . . . . . . . . . .30
3.4.3 混淆矩陣(預測結果) . . . . . . . . . . . . . . . . . . . . . .36
3.5 投票策略 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.6 偽標籤生成方式 . . . . . . . . . . . . . . . . . . . . . . . . . . .40
3.7 增量學習 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.8 模型與資料集選用 . . . . . . . . . . . . . . . . . . . . . . . . 42
Chapter 4 實驗結果與討論 43
4.1 第一階段模型結果 . . . . . . . . . . . . . . . . . . . . . . . . .43
4.2 第二階段模型結果 . . . . . . . . . . . . . . . . . . . . . . . . .47
4.3 第三階段模型結果 . . . . . . . . . . . . . . . . . . . . . . . . .49
4.3.1 模型選擇 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.2 偽標籤生成結果 . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.3 增量學習結果 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Chapter 5 結論與未來展望 54
5.1 結論 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.2 未來展望 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
References 56

[1] 連紹婷, “利用 yolo 辨識黴菌,” 碩士論文, 國立中山大學應用數學系, 2023.
[2] J. Teutschbein, D. Albrecht, M. Pötsch, R. Guthke, V. Aimanianda, C. Clavaud, J.- P. Latgé, A. A. Brakhage, and O. Kniemeyer, “Proteome profiling and functional classification of intracellular proteins from conidia of the human-pathogenic mold aspergillus fumigatus,” Journal of proteome research, vol. 9, no. 7, pp. 3427–3442, 2010.
[3] J. Houbraken, S. Kocsubé, C. M. Visagie, N. Yilmaz, X.-C. Wang, M. Meijer, B. Kraak, V. Hubka, K. Bensch, R. Samson et al., “Classification of aspergillus, penicillium, talaromyces and related genera (eurotiales): An overview of families, genera, subgenera, sections, series and species,” Studies in mycology, vol. 95, pp. 5–169, 2020.
[4] R. K. C. Billones, E. J. Calilung, E. P. Dadios, and N. Santiago, “Image-based macroscopic classification of aspergillus fungi species using convolutional neural networks,” in 2020 IEEE 12th International Conference on Humanoid, Nanotech- nology, Information Technology, Communication and Control, Environment, and Management (HNICEM). IEEE, 2020, pp. 1–4.
[5] ——, “Aspergillus species fungi identification using microscopic scale images,” in
2020 IEEE 12th International Conference on Humanoid, Nanotechnology, Infor- mation Technology, Communication and Control, Environment, and Management (HNICEM). IEEE, 2020, pp. 1–5.
[6] N. R. R. b. M. Radzuan, H. B. Jaafar, and A. S. B. A. Nasir, “An identification of aspergillus species: A comparison on supervised classification methods,” in Pro- ceedings of the 12th National Technical Seminar on Unmanned System Technology 2020: NUSYS’20. Springer, 2022, pp. 965–977.
[7] K. Diba, P. Kordbacheh, S. Mirhendi, S. Rezaie, and M. Mahmoudi, “Identifica- tion of aspergillus species using morphological characteristics,” Pakistan journal of medical sciences, vol. 23, no. 6, p. 867, 2007.
[8] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “Yolov4: Optimal speed and ac- curacy of object detection,” arXiv preprint arXiv:2004.10934, 2020.
[9] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6023–6032.
[10] Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,” arXiv preprint arXiv:2107.08430, 2021.
[11] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in Proceedings of the IEEE conference on computer vi- sion and pattern recognition, 2016, pp. 779–788.
[12] G. Jocher, “YOLOv5 release v7.0,” https://github.com/ultralytics/yolov5/releases/ tag/v7.0, 2022.
[13] G. Jocher, “YOLOv8 release v8.1.0,” https://github.com/ultralytics/ultralytics/releases/ tag/v8.1.0, 2024.
[14] V. Dumoulin and F. Visin, “A guide to convolution arithmetic for deep learning,” arXiv preprint arXiv:1603.07285, 2016.
[15] K.He,X.Zhang,S.Ren,andJ.Sun,“Spatialpyramidpoolingindeepconvolutional networks for visual recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 37, no. 9, pp. 1904–1916, 2015.
[16] C.-Y. Wang, I.-H. Yeh, and H.-Y. M. Liao, “Yolov9: Learning what you want to learn using programmable gradient information,” arXiv preprint arXiv:2402.13616, 2024.
[17] 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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7464–7475.
[18] C.-Y. Wang, H.-Y. M. Liao, Y.-H. Wu, P.-Y. Chen, J.-W. Hsieh, and I.-H. Yeh, “Csp- net: A new backbone that can enhance learning capability of cnn,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 390–391.