I am working on Diabetic Retinopathy Dataset. The Size of images is realtively large. I am currently using DenseNet169 Architecture for making a 5-class classifier and getting approximately 70% accuracy.
However DenseNet takes 112 x 112 Inputs and I feel the image are having minute details which is not getting captured. Can you suggest some architectures which takes bigger inputs (512*512) which will work well in identifying small differences in images.
Alternatively can i add Conv Blocks before the DenseNet169. If so can you guide me how to do it efficiently?