在過去幾年中，深度學習開始在不同領域的醫療保健中產生巨大影響。深度學習方 法在醫療保健領域比較常見的應用在於設計一個可以輔助疾病診斷和自動分析醫學 圖像的系統，用以幫助制定治療計劃。人眼對於醫學圖像辨識的難度相當高，即便深 度學習 (DL) 方法在圖像識別方面表現良好，應用在醫學影像中仍是前所未有的挑戰。 在虹膜圖像處理中實施電腦輔助技術，並將虹膜學與中醫 (TCM) 相結合是數位圖像 處理和人工智慧研究的一個具有挑戰性的領域。本論文重點將討論如何處理虹膜診 斷中的挑戰性問題：(1) 如何開發基於深度學習的計算機輔助診斷 (CAD) 方法來自 動化虹膜學應用程序； (2) 如何處理數據集中的類別不平衡問題；(3) 如何將圖像分 辨率提高使得能夠在後期使用深度學習技術。因此，訓練深度學習模型以識別特定 模式是一項艱鉅的任務。 對於第一個問題，本篇提出的方法結合了基於虹膜識別框架的電腦視覺技術和使用 卷積神經網路的圖像分類方法，替為醫療保健行業中創造了一種新方法。 數據集當中存在戴眼鏡的眼睛圖像、瞳孔過大和過小的圖像、虹膜位置錯位的圖像 等異常類別，造成數據集類別高度不平衡。 這種異常情況會引起虹膜分割和遮罩預 估的失敗，進而導致虹膜識別和虹膜診斷的失敗。為了解決類別不平衡問題並生成 更多稀有虹膜圖像，我們提出了一種數據增強方法，該方法使用具有梯度懲罰的條 件式 Wasserstein 生成對抗網路（CWGAN-GP）生成少數虹膜圖像，從而為稀有數據 收集節省了大量人力成本。 在數位影像中，圖像分辨率在各種影像處理技術皆為重要因素。若分辨率低，則難以 被虹膜學與虹膜辨識使用。為了提高圖像分辨率來獲得更好的分類效果，我們提出單 張圖像超分辨率(SISR)演算法─DDA-SRGAN，基於生成對抗式網路(GAN)中使用掩碼 注意機制(mask-attention mechanism)。;In the past few years, deep learning (DL) has emerged to give big impacts in different areas of healthcare. The application of deep learning approaches in healthcare aims to design a system that can assist diagnosis of diseases and automate the analysis of medical images to help treatment planning. DL methods perform adequately in image recognition, nevertheless medical images show unprecedented challenges. Implementing computer-aided techniques in iris image processing and connecting iridology with Traditional Chinese Medicine (TCM) is a challenging area of research in digital image processing and artificial intelligence. This thesis focuses on how to deal with the challenging problems in iridology: (1) how to develop a DL-based Computer-Aided Diagnosis (CAD) methodology to automate the iridology applications; (2) how to deal with the problem of class imbalance in dataset, (3) how to enhance the image resolution to a scale which enables the deep learning technique in later stage. Consequently, it causes a daunting task to train a deep learning model to recognize specific patterns. For the first problem, the proposed work combined the computer vision techniques based on the iris recognition framework and image classification approaches using convolutional neural networks to make a new approach in the healthcare profession. The dataset is highly imbalanced due to scarcity of the abnormal classes such as images of eyes with glasses, oversized and undersized pupils, and misaligned iris locations … etc. Such abnormal cases will cause the failure of iris segmentation and mask estimation, which will lead to the failure of iris recognition as well as iridology classification. To address the class imbalance problem and generate more rare cases of iris, we propose a data augmentation method that uses the Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) to generate minority iris images, which saves extensive labor costs for rare data collection. In digital imaging, image resolution is a primary factor for the progress of various image processing technologies. If the resolution is low, then it is hard to perform iridology and iris recognition. In order to enhance the image resolution for a better classification rate, we propose DDA-SRGAN, a single image super-resolution (SISR) algorithm, based on the mask-attention mechanism used in Generative Adversarial Network (GAN).
