1. Iffy AI acceptance. Chiwome and colleagues note that radiology has long been a technology-driven specialty. However, its not just radiologists who need to buy in to AIs role in their work.

There is a need to sensitize [referring physicians] about AI through different channels to make the adoption of AI smooth, the authors write. We also need consent from patients to use AI on image interpretation. Patients should be able to choose between AI and humans.

2. The commonness of insufficient training data. No matter how massive the inputs, image-based training datasets arent enough if the data isnt properly labeled for the training, the authors point out. Image labeling takes a lot of time and needs a lot of effort, and also, this process must be very robust, they write.

Also in this category of challenges is the inescapability of rare conditions. Not only are highly unusual findings too few and far between to train algorithms, Chiwome and co-authors write, but nonhuman modes of detection sometimes also mistake image noise and variations for pathologies.

Along those same lines, if image data used in training is from a different ethnic group, age group or different gender, it may give different results if given raw data from other diverse groups of people.

Excerpt from:

Does AI mean the end for breast radiologists? - AI in Healthcare