The authors of this study reviewed research on AI algorithms relating to computed tomography (CT) of the head in order to verify to what degree it is true that AI software for applications in radiology must be transferable to other real-world problems. It was discovered that current research on AI for head CT is rarely reproducible, does not match with real-world epidemiology, and should be more transparent. Because AI in radiology is experiencing rapid growth, it must meet scientific standards.
Key points
- Most research on machine learning for head CT imaging is not reproducible.
- Algorithms are not open source in most cases.
- Balancing the training data rarely mirrors real-world epidemiology.
- Graphical illustrations of model architecture were designed heterogeneously.
Authors: Felix Gunzer, Michael Jantscher, Eva M. Hassler, Thomas Kau & Gernot Reishofer
