| Summary: | 碩士 === 國立交通大學 === 統計學研究所 === 107 === Age-related macular degeneration (AMD) is a disease that affects many elderly populations and causing severe loss of vision to the sufferer. Polypoidal choroidal vasculopathy (PCV) is a retinal disease that involves blood vessels that are characterized by the presence of polypoidal lesions, this disease appears commonly in Asian 's people. AMD and PCV are two different diseases, but share fundamental similarities so that misclassification often occurs.
This study classifies AMD and PCV color fundus images collected in Taipei Veterans General Hospital. We found much more AMD data than PCV data in the medical images. This study used three direction methods, including resampling method, synthetic method, and cost-sensitive learning to explore the effect of the imbalanced data on the classification performance of deep learning and compare the results of distinct methods. The performance of imbalanced data calculated by the Matthews correlation coefficient (MCC)
The result shows that cost-sensitive learning has the best performance, which has the MCC value of 88.11%. The performance of the best model evaluated by using color fundus images from Thailand. Besides classification, this study also localizes the important regions in the image for classification which can help a clinician in quantitatively observing the deterioration and thereby, aiding in the diagnostic process of AMD and PCV images using the gradient weighted class activation mapping (Grad-CAM).
|
|---|
