- 學位論文
使用生成對抗網路擴增口腔癌病徵照片之研究
A Study on Augmenting Medical Images of Oral Cancer Using Generative Adversarial Networks
本文將於2028/03/18開放下載。若您希望在開放下載時收到通知，可將文章加入收藏。
摘要
在醫療影像分析中，深度學習模型通常依賴大量標記數據來實現卓越的效能。然而，由於醫療影像數據的獲取受到限制，特別是在特定疾病或罕見病例的情況下，小數據集會限制模型的表現。為了解決這一挑戰，本研究提出了一種結合影像合成與生成對抗網絡（Generative Adversarial Networks, GAN）的數據增強方法。該方法生成的醫療影像能顯著提升深度學習模型在醫療影像分割任務中的表現。 口腔癌是一種發生在口腔內部的惡性腫瘤，其中有高達90%的病例源自口腔潛在疾病的惡化。一旦確診為口腔癌，患者的五年存活率將降至僅56%，顯示出早期發現與治療的重要性。然而，由於台灣部分地區醫療人力不足，許多患者無法及時接受檢查。因此，我們選定口腔潛在疾病中盛行率高且外觀容易觀察的口腔白斑作為主要研究目標，以期提高早期診斷與治療的可能性。 我們在網路上蒐集並篩選出有公信力的口腔白斑影像共15張，以及由埔里基督教醫院蒐集的426張正常口腔影像。我們的研究首先通過影像合成技術，將白斑病特徵貼合至正常舌頭影像上，然後利用GAN架構修復合成影像中的潛在異常。隨後，使用傳統數據增強方法與我們生成的人造數據混合在一起進行訓練，並比較深度學習模型的效能。實驗結果顯示，未採用我們的方法時，模型的Intersection over Union（IOU）僅達76%；加入人造數據後，效能顯著提升至80.9%。 這是第一個對口腔白斑正常拍攝影像進行分割的系統，透過將病徵黏貼至正常區域，使得資料擴增達到一定的效果，並且此系統將可以推廣至醫療落後的地區，作為快篩使用。甚至此種創新的數據增強方法可推廣至其他領域，為小數據集環境中的模型效能提升提供有效解決方案。
並列摘要
In medical image analysis, deep learning models typically rely on large amounts of labeled data to achieve outstanding performance. However, the acquisition of medical imaging data is often constrained, particularly in cases of specific diseases or rare conditions, limiting the performance of models trained on small datasets. To address this challenge, this study proposes a data augmentation approach that combines image synthesis with Generative Adversarial Networks (GANs). The medical images generated by this method significantly enhance the performance of deep learning models in medical image segmentation tasks. Oral cancer is a malignant tumor that occurs within the oral cavity, with up to 90% of cases stemming from the progression of oral potentially malignant disorders. Once diagnosed with oral cancer, the five-year survival rate drops to only 56%, highlighting the importance of early detection and treatment. However, due to insufficient medical resources in certain regions of Taiwan, many patients are unable to receive timely examinations. Therefore, we have chosen oral leukoplakia, a common and visually identifiable oral potentially malignant disorder, as the primary focus of our research, aiming to enhance the likelihood of early diagnosis and treatment. We collected and selected 15 credible images of oral leukoplakia from online sources and 426 normal oral images from Puli Christian Hospital. Our study first employed image synthesis techniques to overlay leukoplakia features onto normal tongue images. Then, we used a GAN architecture to correct potential anomalies in the synthesized images. Subsequently, we combined traditional data augmentation methods with our generated synthetic data for training and compared the performance of deep learning models. The experimental results showed that without our approach, the model's Intersection over Union (IoU) reached only 76%. However, after incorporating synthetic data, the performance significantly improved to 80.9%. This is the first system designed to segment normal images of oral leukoplakia. By pasting lesions onto normal areas, data augmentation achieves significant results. This system can be promoted in medically underserved regions as a tool for rapid screening. Furthermore, this innovative data augmentation method can be extended to other fields, offering an effective solution for improving model performance in small dataset scenarios.
參考文獻
[1]. Chen, Y., Yang, X. H., Wei, Z., Heidari, A. A., Zheng, N., Li, Z., ... & Guan, Q. (2022). Generative adversarial networks in medical image augmentation: a review. Computers in Biology and Medicine, 144, 105382.
[2]. Xun, S., Li, D., Zhu, H., Chen, M., Wang, J., Li, J., ... & Huang, P. (2022). Generative adversarial networks in medical image segmentation: A review. Computers in biology and medicine, 140, 105063.
[3]. Han, C., Hayashi, H., Rundo, L., Araki, R., Shimoda, W., Muramatsu, S., ... & Nakayama, H. (2018, April). GAN-based synthetic brain MR image generation. In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018) (pp. 734-738). IEEE.
[4]. You, Q., Wan, C., Sun, J., Shen, J., Ye, H., & Yu, Q. (2019, July). Fundus image enhancement method based on CycleGAN. In 2019 41st annual international conference of the IEEE engineering in medicine and biology society (EMBC) (pp. 4500-4503). IEEE.
[5]. . Wagner, S. J., Khalili, N., Sharma, R., Boxberg, M., Marr, C., De Back, W., & Peng, T. (2021). Structure-preserving multi-domain stain color augmentation using style-transfer with disentangled representations. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VIII 24 (pp. 257-266). Springer International Publishing.