**4. How to generate AI in musculoskeletal medicine?**

To generate an AI, a large volume of labeled data are needed to train the AI system and generate a robust algorithm. Otherwise, the algorithm can only be applied in a limited way. In the case of radiology, 49% of the papers using DL use databases of 101 to 1000 cases, 25% less than 100 cases, and only 6% use more than 10,000 cases [7]. It seems necessary that centers could coordinate to increase the size of their databases. In this regard, there are de-identified public databases that can be used to train AI algorithms, such as musculoskeletal radiographs (MURA), with almost 41,000 images of the upper extremity labeled as fracture or nonfracture by radiologists [8].

Many times, images processed by AI systems are manually selected, which is very time-consuming. It is vital that the database that is going to train the AI is appropriate to what is to be analyzed and has no flaws. In addition, it is recommended for the data to be homogeneous and of a volume proportional to the complexity of the computational task.

Unsupervised learning is likely to be critical in the future for building new AI systems. However, most successful AIs currently use supervised learning that may actually hinder their development [9].

One element used in some algorithms is heat maps in DL systems. Their use allows us to find out the part of the image that contributes the most within the analysis and reduce the impact of incorrect data. For example, if the heat map points out that a part of the image is being analyzed while the lesion is in a different one, it can be discovered that the algorithm is not processing the correct data.
