Can AI differentiate a circle from a square?

It seems that powerful deep learning techniques should breeze through a problem of differentiating a circle from a square, two simple geometric shapes. Perhaps, the problem is so trivial that only amateurs would ponder about such a topic. This may explain why top search results on this topic only lead to a few discussions on stackexchange type boards instead of scientific articles. This is an important question from radiology researcher’s perspective. Shape is only one of thousands of features that humans take into account when evaluating images, but it is an important one. For example, “irregularly shaped erosion” and “smoothly shaped round cyst” describe two very different concepts. Here is an attempt by radiologist without any formal computer science training other than online MOOCs to answer this question and learn something about AI in the process.

How accurate is AI in differentiating simple geometric shapes?