隨著深度學習技術在醫學影像分析領域的迅速發展，準確分割不同器官和病灶對於診斷和治療至關重要。然而，由於醫學影像數據的獲取和標註困難，常常存在不平衡的情況，即某些類別的樣本數量明顯少於其他類別，這給分割模型的訓練帶來了挑戰。尤其是對於頭頸部等區域，由於器官體積較小，即使經過標註，相關病灶或結構在整個影像中所佔比例也很有限，加劇了類別不平衡問題。

本研究旨在探索頭頸部腦幹和脊髓器官分割的最佳模式。我們首先採用了不同的直方圖均衡化技術，分別為：直方圖均衡化技術（Histogram Equalization, HE）、對比度受限自適應直方圖均衡化（Contrast Limited Adaptive Histogram Equalization, CLAHE）以及伽瑪校正（gamma correction）對原始影像進行增強預處理，以提高影像對比度和細節可視化的效果。接著，我們在腦幹（brainstem）和脊髓（spinal cord）這兩個頭頸部器官的分割任務上，分別測試了U-Net和LinkNet模型。

實驗結果表明，針對腦幹的分割，使用U-Net搭配混合dice 損失函數並對影像進行gamma correction前處理，可以達到最佳的結果，其DSC為0.861607。針對脊髓的分割，使用U-Net搭配mixed損失函數並對影像進行CLAHE前處理，可以取得最佳的分割結果，其DSC為0.839489。針對腦幹與脊髓的電腦斷層影像分割，建議針對個別器官進行獨立的模型訓練，再進行器官分割會有較佳的結果，並選擇合適的直方圖均衡化技術可以有效提升影像的特徵，而提升影像分割結果。

未來的研究可以著重於探索這些前處理技術和損失函數在多器官分割中的最佳組合，並進一步探討如何在高解析度影像上應用這些技術，同時也期待在未來的實驗中，能夠探索如何在不影響模型訓練效率的情況下，保留影像的高分辨率特徵，以充分利用注意力機制的優勢，將深度學習廣泛結合到臨床應用當中。

With the rapid development of deep learning technology in medical image analysis, accurately segmenting different organs and lesions is crucial for diagnosis and treatment. However, due to the difficulty in acquiring and annotating medical image data, there is often an imbalance in the data, where certain categories have significantly fewer samples than others. This poses challenges for training segmentation models. Especially for areas such as the head and neck, the small volume of organs means that even when annotated, the relevant lesions or structures occupy a very limited proportion of the entire image, exacerbating the class imbalance problem.

This study explores the optimal methods for segmenting the brainstem and spinal cord in head and neck organ-at-risk images. We first employed different histogram equalization techniques, including Histogram Equalization (HE), Contrast Limited Adaptive Histogram Equalization (CLAHE), and gamma correction, to enhance the original images, improving contrast and detail visualization. Next, we tested different deep learning segmentation models, specifically U-Net and LinkNet, on the segmentation tasks of the brainstem and spinal cord.

The experimental results indicate that for the segmentation of the brainstem, using U-Net with a mixed loss function and gamma correction preprocessing, or using LinkNet with Dice loss and gamma correction preprocessing, can achieve better results. For the segmentation of the spinal cord, using U-Net with Dice loss and CLAHE preprocessing, or using LinkNet with dice loss or mixed loss and HE preprocessing, can yield better outcomes. The results suggest that training independent models for each small organ separately can lead to better segmentation results for the segmentation of CT images of the brainstem and spinal cord. Histogram equalization techniques effectively enhance image features, thereby improving segmentation outcomes.

Future research can focus on exploring the optimal combination of these preprocessing techniques and loss functions in multi-organ segmentation, and further investigate how to apply these techniques to high-resolution images. Additionally, future experiments aim to explore how to preserve high-resolution features without compromising model training efficiency, fully leveraging the advantages of attention mechanisms and widely integrating deep learning into clinical applications.

[1] Wang, Q., et al. (2021). Deep Learning-Augmented Head and Neck Organs at Risk Segmentation From CT Volumes. *Frontiers in Oncology*.
[2] Simon, A., et al. (2022). Deep Learning-Based Segmentation of Head and Neck Organs-at-Risk with Clinical Partially Labeled Data. *Entropy*.
[3] Wang, X., et al. (2023). Automatic Segmentation of Organs at Risk in Head and Neck Cancer Patients from CT and MRI Scans.
[4] Men, K., et al. (2019). Deep Learning to Achieve Clinically Applicable Segmentation of Head and Neck Anatomy for Radiotherapy.
[5] Soomro, M. H., et al. (2021). OARnet: Automated Organs-at-Risk Delineation in Head and Neck CT Images.
[6] Morgan, H.E., Sher, D.J. (2020) Adaptive radiotherapy for head and neck cancer. Cancers Head Neck 5, 1.
[7] LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE.
[8] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems.
[9] Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
[10] Szegedy, C., et al. (2015). Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[11] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[12] LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE.
[13] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems.
[14] Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
[15] Szegedy, C., et al. (2015). Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[16] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[17] MiHElucci, U. (2022). Convolutional Neural Networks. In: Applied Deep Learning with TensorFlow 2. Apress, Berkeley, CA.
[18] Beysolow II, T. (2017). Convolutional Neural Networks (CNNs). In: Introduction to Deep Learning Using R. Apress, Berkeley, CA.
[19] Alzubaidi, L., Zhang, J., Humaidi, A.J. (2021) et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data 8, 53. https://doi.org/10.1186/s40537-021-00444-8
[20] Cong, S., Zhou, (2023) Y. A review of convolutional neural network architectures and their optimizations. Artif Intell Rev 56, 1905–1969
[21] Z. Li, F. Liu, W. Yang, S. Peng and J. Zhou, (2022) "A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects," in IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 6999-7019
[22] Ronneberger, O., FisHEr, P., Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab, N., Hornegger, J., Wells, W., Frangi, A. (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science(), vol 9351. Springer, Cham.
[23] A. Chaurasia and E. Culurciello (2017), "LinkNet: Exploiting encoder representations for efficient semantic segmentation," 2017 IEEE Visual Communications and Image Processing (VCIP), St. Petersburg, FL, US, pp. 1-4
[24] Brett Clark, Nicholas Hardcastle, Leigh A. Johnston, James Korte (2024), Transfer learning for auto‐segmentation of 17 organs‐at‐risk in the head and neck: Bridging the gap between institutional and public datasets, Medical Physics, 10.1002/mp.16997, 51, 7, (4767-4777)
[25] Wenhui Lei, HaoHEn Mei, Zhengwentai Sun, Shan Ye, Ran Gu, Huan Wang, Rui Huang, Shichuan Zhang, Shaoting Zhang, Guotai Wang (2021), Automatic segmentation of organs-at-risk from head-and-neck CT using separable convolutional neural network with hard-region-weighted loss, Neurocomputing, Volume 442, Pages 184-199,ISSN 0925-2312
[26] Z. Li, F. Liu, W. Yang, S. Peng and J. Zhou (2022). A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. in IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 6999-7019, Dec. 2022, doi: 10.1109/TNNLS.2021.3084827.
[27] Zou, K. H., Warfield, S. K., Bharatha, A., Tempany, C. M., Kaus, M. R., Haker, S. J., Wells, W. M., 3rd, Jolesz, F. A., & Kikinis, R. (2004). Statistical validation of image segmentation quality based on a spatial overlap index. Academic radiology, 11(2), 178–189.
[28] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. (2014). Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1 (January 2014), 1929–1958.
[29] Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2012). Improving neural networks by preventing co-adaptation of feature detectors.
[30] Sergey Ioffe and Christian Szegedy. 2015. Batch normalization: accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37 (ICML'15). JMLR.org, 448–456.
[31] Bjorck, J., Gomes, C.P., & Selman, B. (2018). Understanding Batch Normalization. Neural Information Processing Systems.
[32] Ouyang, Y., et al. (2004). Thresholding-based image segmentation in medical imaging. Journal of Medical Imaging and Health Informatics.
[33] Ruan, S., et al. (2009). Edge detection techniques in head and neck tumor segmentation. IEEE Transactions on Medical Imaging.
[34] Leung, S., et al. (2006). Region growing algorithms in medical image segmentation. Computers in Biology and Medicine.
[35] Anquez, J., et al. (2011). Watershed-based segmentation for head and neck radiotherapy planning. Physics in Medicine & Biology.