近年來，口腔癌在台灣癌症十大死因中位居第五名，而早期口腔癌的五年存活率可以高達77.4%，而在全球每年死於口腔癌的人數更是超過30萬人，為了在早期發現口病變，許多口腔癌篩檢的方式被提出，其中光學快速且非侵入式之方法最適合用於初步篩檢。其原理為利用口腔細胞自體螢光和型態上的變化來做為判別的依據。在先前的研究中我們已開發出用於拍攝口腔自體螢光影像之儀器系統，並應用於臨床資料收集及測試，此系統將所拍攝的螢光影像經由醫生圈選興趣區域，在該區域擷取紋理特徵並做分類已有良好的分類正確率。然而，由醫生圈選興趣區域需要專業性且耗時的工作。本研究希望能透過深度學習中的深度卷積神經網路來達到自動化圈選興趣區域的效果，如此一來就能夠提供更快速且有效地用螢光影像來進行口腔癌篩檢。
由於口腔癌影像沒有固定型態和特徵，為了達到上述目的，本研究透過紋理影像擷取的前處理來產生固定特徵的影像，我們分別使用了小波轉換和賈伯濾波器來做為我們的紋理影像擷取方法，接著計算紋理影像的標準差，並將所產生的結果輸入到深度卷積神經網路模型，此外我們也將癌症偵測和興趣區域的圈選整合為一個神經網路模型，使癌症偵測和興趣區域可以在相同模型中完成以提升系統的效率。在我們神經網路模型中包含兩個分支，在第一條分支我們分別利用剩餘網路(residual network)和初始模組(inception module)架構來做為我們神經網路模型中用於偵測癌症區域，在第二條分支中我們分別利用全卷積網路(fully convolution network)和特徵金字塔網路(feature pyramid network)架構用於興趣區域的圈選。由實驗結果發現基於賈伯濾波器的紋理影像比基於小波轉換的紋理影響提供更好的結果，而初始模組(inception module)又比剩餘網路(residual network)有更高的特異性，最後我們也比較傳統的卷積神經網路架構和沒有圈選興趣區域分支的神經網路架構來驗證我們的神經網路模型，而在臨床上我們為針對靈敏性和特異性做權衡，以降低一些特異性來達到高靈敏性的結果，並藉由口腔癌的機率分佈圖和癌前病變偵測來進行臨床上的分析，藉此來達到適用於臨床的系統來及早發現口腔癌並防止其惡化。

Oral cancer consistently ranks as one of the top five leading causes of cancer death in Taiwan. And over 300 thousand people dead caused by oral cancer every year. However, the five-year survival rate of early-stage oral cancer patients is as much as 77.4%. Thus, it is highly desirable to detect oral cancer at the earliest stage possible. Optical imaging techniques provide an efficient and feasible approach for the preliminary diagnosis of oral cancer. For example, in a previous study by the present group, an oral cavity auto-fluorescence imaging system was developed for the acquisition of clinical images, and an algorithm was then proposed for analyzing the image features within the region of interest (ROI). However, the algorithm required the ROI to be manually marked by a doctor or healthcare practitioner. As a result, the diagnosis procedure was expensive and time-consuming. Accordingly, the present thesis proposes a deep convolutional neural network for detecting cancerous regions within the input image and then marking the ROI automatically.

In the proposed model, two different feature extraction methods, namely wavelet transformation, and the Gabor filter are employed to extract the texture features from the input image. A sliding window is then applied to compute the standard deviation values of the texture image. Finally, the standard deviation values are used to construct a feature map, which is partitioned into multiple patches and used as the input data to the deep convolutional network model. The model contains two branches, namely an upper branch based on a residual network or inception module to perform oral cancer detection, and a lower branch with a fully convolutional network or feature pyramid network configuration to perform semantic segmentation and ROI marking. The experimental results show that the optimal cancer detection and segmentation performance is obtained using low-frequency Gabor filter features and the inception module configuration.

An investigation is performed on the relationship between the sensitivity and specificity of the proposed framework. It is shown that the sensitivity of the detection results can be enhanced by easing the criterion employed in the classification branch of the framework to identify cancerous regions. Using the proposed tradeoff solution, cancer probability output maps are produced for subsequent clinical analysis. Finally, the applicability of the proposed model to pre-cancer screening is investigated. It is shown that the model has significant potential for the detection of oral cancer at an early stage.

1. Health Promotion Administration, Ministry of Health and Welfare, Taiwan. Cancer Prevention and Control. Available at:
http://www.hpa.gov.tw/BHPNet/English/ClassShow.aspx?No=201401280006
2. Farah, Camile S., et al. Advances in Early Detection and Diagnostic Adjuncts in Oral Cavity Cancer. Contemporary Oral Oncology. Springer International Publishing, 2017. 355-421.
3. Chiang, C.-J., et al., Incidence and survival of adult cancer patients in Taiwan, 2002-2012. Journal of the Formosan Medical Association (2015), 1–13.
4. Thankappan, Krishnakumar, and Moni Abraham Kuriakose. Sentinel Node Biopsy in Oral Cancer. Contemporary Oral Oncology. Springer International Publishing, 2017. 211-233.
5. Remmerbach, T. W., et al. Liquid-based versus conventional cytology of oral brush biopsies: a split-sample pilot study. Clinical Oral Investigations (2017): 1-6.
6. Wang, Xiaoqian, Karolina Elżbieta Kaczor-Urbanowicz, and David TW Wong. Salivary biomarkers in cancer detection. Medical Oncology 34.1 (2017): 7.
7. Chen, Ya-Wei, et al. Use of methylene blue as a diagnostic aid in early detection of oral cancer and precancerous lesions. British Journal of Oral and Maxillofacial Surgery 45.7 (2007): 590-591.
8. Schwarz, Richard A., et al. Autofluorescence and diffuse reflectance spectroscopy of oral epithelial tissue using a depth-sensitive fiber-optic probe. Applied optics 47.6 (2008): 825-834.
9. De Veld, Diana CG, et al. Autofluorescence and diffuse reflectance spectroscopy for oral oncology. Lasers in surgery and medicine 36.5 (2005): 356-364.
10. De Veld, Diana CG, et al. Clinical study for classification of benign, dysplastic, and malignant oral lesions using autofluorescence spectroscopy. Journal of biomedical optics 9.5 (2004): 940-950.
11. Hung, Hsin-Ru. Application of Autofluorescence Imaging on Oral Cancer Screening. (2013): 1-45.
12. L. Z. Li, et al., Quantitative Magnetic Resonance and Optical Imaging Biomarkers of Melanoma Metastatic Potential. Proceedings of the National Academy of Sciences (Impact Factor: 9.67). 05/2009; 106(16):6608-13.
13. Cicciù, Marco, et al. Tissue Fluorescence Imaging (VELscope) for Quick Non-Invasive Diagnosis in Oral Pathology. Journal of Craniofacial Surgery (2017).
14. Huang, Teng-Yi. Application of Autofluorescence Imaging On Clinical Oral Cancer Screening. (2014): 1-56.
15. Chang, Ming-Hsuan. Texture Feature Analysis for Oral Cancer Detection. (2016): 1-58.
16. Iakovidis, Dimitris K., Eystratios G. Keramidas, and Dimitris Maroulis. Fuzzy local binary patterns for ultrasound texture characterization. International Conference Image Analysis and Recognition. Springer Berlin Heidelberg, 2008.
17. Kruizinga, Peter, and Nikolay Petkov. Nonlinear operator for oriented texture. IEEE Transactions on image processing 8.10 (1999): 1395-1407.
18. Redmon, Joseph, et al. You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.
19. Noh, Hyeonwoo, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. Proceedings of the IEEE International Conference on Computer Vision. 2015.
20. Ledig, Christian, et al. Photo-realistic single image super-resolution using a generative adversarial network. arXiv preprint arXiv:1609.04802 (2016).
21. He, Kaiming, et al. Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.
22. Szegedy, Christian, et al. Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015.
23. He, Kaiming, et al. Mask R-CNN. arXiv preprint arXiv:1703.06870 (2017).
24. Long, Jonathan, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015.
25. Lin, Tsung-Yi, et al. Feature Pyramid Networks for Object Detection. arXiv preprint arXiv:1612.03144 (2016).
26. Grigorescu, Simona E., Nicolai Petkov, and Peter Kruizinga. Comparison of texture features based on Gabor filters. IEEE Transactions on Image processing 11.10 (2002): 1160-1167.
27. Hamamoto, Yoshihiko, et al. A Gabor filter-based method for recognizing handwritten numerals. Pattern Recognition 31.4 (1998): 395-400.
28. Weldon, Thomas P., William E. Higgins, and Dennis F. Dunn. Efficient Gabor filter design for texture segmentation. Pattern recognition 29.12 (1996): 2005-2015.
29. Porter, R., and N. Canagarajah. Robust rotation-invariant texture classification: wavelet, Gabor filter and GMRF based schemes. IEE Proceedings-Vision, Image and Signal Processing 144.3 (1997): 180-188.
30. S. Arivazhagan, et al., Texture classification using wavelet transform. Pattern Recognition Letters 24 (2003) 1513–1521.
31. K. Huang, et al., Wavelet Feature Selection for Image Classification. IEEE Transactions on Image Processing, Vol. 17, No. 9, September 2008.
32. B. Remeseiro, et al., Automatic classification of the interferential tear film lipid layer using colour texture analysis. Computer Methods and Programs in Biomedicine III (2013) P.93-103
33. Girshick, Ross, et al. Rich feature hierarchies for accurate object detection and semantic segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition. 2014.
34. He, Kaiming, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition. European Conference on Computer Vision. Springer International Publishing, 2014.
35. Girshick, Ross. Fast r-cnn. Proceedings of the IEEE International Conference on Computer Vision. 2015.
36. Ren, Shaoqing, et al. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems. 2015.
37. Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer International Publishing, 2015.
38. Iizuka, Satoshi, Edgar Simo-Serra, and Hiroshi Ishikawa. Let there be color!: joint end-to-end learning of global and local image priors for automatic image colorization with simultaneous classification. ACM Transactions on Graphics (TOG) 35.4 (2016): 110.
39. A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NIPS, 2012.
40. Karen Simonyan, Andrew Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition, ICLR, 2015.
41. Hinton, Geoffrey E., et al. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580 (2012).
42. Kingma, Diederik, and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).
43. Ioffe, Sergey, and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167 (2015).
44. Lin, Min, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400 (2013).