Yapay zeka ile diyabetik retinopati tespiti
Küçük Resim Yok
Tarih
2025
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Bursa Teknik Üniversitesi
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Diyabetik retinopati (DR), diyabetin neden olduğu göz hastalıkları arasında en yaygın ve ciddi olanlardan biridir ve retinadaki kan damarlarında mikroanevrizma, kanama ve eksudalar gibi mikrovasküler anormalliklere yol açar. DR'nin erken teşhisi, tedavi sürecinde büyük bir öneme sahip olup, manuel tanı yöntemlerinin zorlukları göz önüne alındığında, bilgisayar destekli tanı sistemlerinin geliştirilmesi oldukça önem kazanmıştır. Bu çalışmada, DR teşhisinde otomatik görüntü sınıflandırma ve lezyon tespitine yönelik farklı derin öğrenme mimarilerinin performansları değerlendirilmiştir. İlk olarak fundus görüntüleri ön işlemesiz VGG16, VGG19, CNN, Xception, InceptionV3 ve ResNet50 modelleri ile DR sınıflandırmasına tabi tutulmuştur. Daha sonra görüntüler, CLAHE (Contrast Limited Adaptive Histogram Equalization) algoritması ile ön işlemden geçirilerek iyileştirilmiş ve yeniden DR sınıflandırması yapılarak tüm sonuçlar karşılaştırılmıştır. Görüntü ön işleme yöntemi olarak kullanılan CLAHE'nin, özellikle küçük lezyonların daha görünür hale getirilmesine katkı sağladığı ve sınıflandırma performansını belirgin şekilde artırdığı gözlemlenmiştir. Elde edilen sonuçlar arasında, VGG16-CLAHE ve CNN-CLAHE modellerinin %99 doğrulukla en yüksek performansı sağladığı belirlenmiştir. Ek olarak bu çalışmada geliştirilen VGG16-CLAHE tabanlı derin öğrenme modeli, kullanıcı dostu bir ara yüz uygulamasına entegre edilmiş ve gerçek zamanlı DR teşhisi yapabilme imkânı sağlanmıştır. Uygulama, DR'nin beş sınıfını (No DR, Mild NPDR, Moderate NPDR, Severe NPDR, PDR) başarılı bir şekilde ayırt edebilmektedir. Gelecekte daha geniş veri setleri ve farklı derin öğrenme mimarileri üzerinde çalışılarak sistem performansının artırılabileceği öngörülmektedir.
Diabetic retinopathy (DR) is one of the most common and serious eye diseases caused by diabetes, leading to microvascular abnormalities in the retinal blood vessels such as microaneurysms, hemorrhages and exudates. Early diagnosis of DR is of great importance in the treatment process, and given the difficulties of manual diagnostic methods, the development of computer-aided diagnostic systems has become very important. In this study, we evaluate the performance of different deep learning architectures for automatic image classification and lesion detection in DR diagnosis. First, fundus images were subjected to DR classification with VGG16, VGG19, CNN, Xception, InceptionV3 and ResNet50 models without preprocessing. Then, the images were preprocessed and enhanced with the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm and DR classification was performed again and all results were compared. It was observed that CLAHE, which was used as an image preprocessing method, contributed to making small lesions more visible and significantly improved the classification performance. Among the results obtained, VGG16-CLAHE and CNN-CLAHE models were found to provide the highest performance with 99% accuracy. In addition, the VGG16-CLAHE-based deep learning model developed in this study was integrated into a user-friendly interface application, enabling real-time DR diagnosis. The application can successfully distinguish five classes of DR (No DR, Mild NPDR, Moderate NPDR, Severe NPDR, PDR). It is envisaged that system performance can be improved by working on larger data sets and different deep learning architectures in the future.
Diabetic retinopathy (DR) is one of the most common and serious eye diseases caused by diabetes, leading to microvascular abnormalities in the retinal blood vessels such as microaneurysms, hemorrhages and exudates. Early diagnosis of DR is of great importance in the treatment process, and given the difficulties of manual diagnostic methods, the development of computer-aided diagnostic systems has become very important. In this study, we evaluate the performance of different deep learning architectures for automatic image classification and lesion detection in DR diagnosis. First, fundus images were subjected to DR classification with VGG16, VGG19, CNN, Xception, InceptionV3 and ResNet50 models without preprocessing. Then, the images were preprocessed and enhanced with the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm and DR classification was performed again and all results were compared. It was observed that CLAHE, which was used as an image preprocessing method, contributed to making small lesions more visible and significantly improved the classification performance. Among the results obtained, VGG16-CLAHE and CNN-CLAHE models were found to provide the highest performance with 99% accuracy. In addition, the VGG16-CLAHE-based deep learning model developed in this study was integrated into a user-friendly interface application, enabling real-time DR diagnosis. The application can successfully distinguish five classes of DR (No DR, Mild NPDR, Moderate NPDR, Severe NPDR, PDR). It is envisaged that system performance can be improved by working on larger data sets and different deep learning architectures in the future.
Açıklama
Anahtar Kelimeler
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
