One of the good things with DL is that it extracts features without our intervention. The conventional ML algorithms need features to be extracted before using them.

At the end, it depends on your project and dataset

Another paper on the matter

One more

Marianne, Thanks for the link.

If you apply transfer learning on an already trained deep model on large database of images, you probably will have better and automated results using deep learning:

https://github.com/arundasan91/Diabetic-Retinopathy