Evaluating Transfer Learning for Macular Fluid Detection with Limited Data

The ability to transfer knowledge learned with large datasets to train classifiers where labeled data is scarce, has made transfer learning a prime step on deep learning applications. The availability of models pre-trained on large datasets makes possible to apply deep learning to a wide variety of computer aided diagnosis tasks including retinal image processing. However, with dozens of deep convolutional network architectures, choosing the right model for transfer learning is not a trivial task. While current art implies that the deeper the model the better, here we verified this assertion by fine-tuning a suite of deep convolutional network architectures on a scenario with limited resources. Contrary to what was expected, deeper architectures did not perform as well with limited data as with large datasets.