Abstract: From our experience on automatic chest X-ray diagnosis with deep learning, we identified successful decisions that could serve as advice for health informatics teams in developing countries that wish to introduce AI in their medical imaging processes. We discuss the use of public datasets and present results of a clinical validation on local images, obtaining area under the Receiver Operator Characteristics curve of 0.82 and 0.94 for lung opacities and pneumothorax detection respectively.

Describe the new knowledge and additional skills the participant will gain after attending your presentation.: The attendee is invited to use our experience as a guide for introducing AI in medical image analysis, particularly in developing countries health institutions. The main lessons learnt are: 1- The quality of the DL algorithm depends on the quality of the labeled dataset, which should be aligned with the intended clinical use. 2- Building a local labeled dataset for clinical validation is a resource-consuming task but it is essential to obtain reliable results. 3- A multidisciplinary team of professionals is key to successfully carry out a continuous improvement process.

Authors:

Facundo Diaz, Hospital Italiano de Buenos Aires
Candelaria Mosquera (Presenter)
Hospital Italiano de Buenos Aires

Alejandro Beresñak, Hospital Italiano de Buenos Aires
Diego Rabinovich, Hospital Italiano de Buenos Aires
Sonia Benitez, Hospital Italiano de Buenos Aires