Side note.

As mentioned before, accuracy of a deep learning model can be impacted by the reference standard. In other words, if a model performs less accurately on a test set, it may simply be a less rigorously defined reference standard (and not the model’s performance) that reduced accuracy. For instance, a model trained on world expert grading of prostate cancer may perform “less well” on a random sample of prostate biopsies from general pathologists for the simple reason that they model may have gotten something “right” that the pathologist got “wrong.” On the other hand, a model can overfit the training data—meaning it is too specific to the data it was trained on and underperforms on external (perhaps more real-world) data. TBL: It takes real intelligence to appropriately assess the reliability of emerging AI in oncology. | QuadShot Team 2020