- 學位論文
以深度學習及影像處理技術應用於糖尿病視網膜病變之不平衡樣本分級
Imbalanced Diabetic Retinopathy Data Classification using Deep Learning and Image Processing Technologies
摘要
糖尿病視網膜病變的檢測需要專業的醫療人士判斷，而傳統的人工檢測有耗時和耗人力等問題，為了解決人工檢測的缺點，近年也有許多研究提出自動化檢測方法。本論文著重在研究深度學習和影像處理方法上的應用，並克服糖尿病視網膜病變資料集之數據不平衡特性而導致分類模型效能下降的問題，訓練一個可供臨床醫師或是醫療團隊有效率的診斷糖尿病視網膜病變的深度學習模型。 本研究將糖尿病視網膜病變之眼底照片使用影像增強、亮度校正、對比度調整等影像處理方法作為數據集前處理步驟，應用一個基於色彩空間轉換、CLAHE、Retinex影像增強算法MSRCR（Multi-Scale Retinex with Color Restoration）、伽瑪校正等方法之融合技術，用於突顯視網膜病理特徵。再使用ResNet50v2、DenseNet121、InceptionV3、Xception、MobileNetV2、InceptionResNetV2等深度學習模型進行訓練後，將糖尿病視網膜病變的數據集依據疾病嚴重程度分成五個級別：無視網膜病變（No DR）、輕度非增殖性視網膜病變（Mild DR）、中等非增殖性視網膜病變（Moderate DR）、嚴重非增殖性視網膜病變（Severe DR）、增殖性視網膜病變（Proliferative DR）。最後比較不同影像預處理方法和深度學習模型的搭配組合之系統性能評估指標。 本研究使用的影像增強之融合技術最終在APTOS 2019數據集上得到的最佳靈敏度（Sensitivity）及特異度（Specificity）的評分穩定落在0.97 ~ 0.99區間內，二次加權Kappa係數（Quadratic Weighted Kappa）為0.897。與其他研究相比也證明了本研究使用的影像增強之融合技術在糖尿病視網膜分級任務上有著優良的系統性能，且MSRCR轉換方法在該影像增強之融合技術架構中是跟其它相關研究不同的貢獻。也證實了透過類別權重方法使模型即使是在數據分佈失衡的情況下可以精準的區分不同嚴重程度視網膜病變特徵。
並列摘要
The detection of diabetic retinopathy requires the judgment of medical professionals, and traditional manual detection has time-consuming and manpower-consuming problems. In order to solve the shortcomings of manual detection, many studies have proposed automatic detection methods for diabetic retinopathy in recent years. This thesis focuses on the application of deep learning and image processing methods, and overcomes the problem of the performance degradation of the classification model caused by imbalanced diabetic retinopathy dataset. Trains an efficient deep learning model for clinicians or medical teams to efficiently diagnose diabetic retinopathy. In this study, image processing methods such as image enhancement, brightness correction, and contrast adjustment use for the pre-processing steps of the datasets for fundus images of diabetic retinopathy, and a fusion technology based on color space conversion, Contrast Limited Adaptive Histogram Equalization, Multi-Scale Retinex with Color Restoration, and gamma correction is applied to highlighting Retinal pathological features. After training with deep learning models such as ResNet50v2, DenseNet121, InceptionV3, Xception, MobileNetV2, InceptionResNetV2, the dataset of diabetic retinopathy is divided into five levels according to the severity of the disease: no retinopathy（No DR）, mild non-proliferative Retinopathy（Mild DR）, moderate non-proliferative retinopathy（Moderate DR）, severe non-proliferative retinopathy（Severe DR）, proliferative retinopathy（Proliferative DR）. Finally, compare the performance evaluation indicators of different image preprocessing methods and deep learning models. The image enhancement fusion technology proposed in this study finally obtained the best sensitivity and specificity scores on the APTOS 2019 data set, and the Quadratic Weighted Kappa coefficient is 0.897. Compared with other studies, it also proves that the method proposed in this study has excellent performance on the diabetic retinal grading task. Furthermore, the utilization of the MSRCR conversion method represents a unique contribution in the image enhancement fusion technology framework, setting it apart from other related research. It is also confirmed that through the class weights techniques, the model can accurately distinguish the features of retinopathy with different severities even in the case of imbalanced data distribution.