The first app for selecting the most effective AI models for treatment of an individual organ

A team from the Department of Electronics, Information and Bioengineering of the Politecnico di Milano, led by dr. Andrea Moglia, has developed the first online application that helps us understand which Artificial Intelligence model is best suited to create 3D images of every individual organ. This makes treatment of the patient more accurate and reliable.

The tool had its genesis in a study published in the important scientific journal Information Fusion, which looked into both generalist and specific AI models. It is designed for health professionals, for technicians who have to create images of organs, lesions or fractures, and for doctors who need to interpret them in order to operate or prescribe treatment.  As Andrea Moglia explained: "With this tool, selecting which models to use for producing the images needed for diagnosis and treatment becomes a far more efficient process. Professionals no longer need to make several attempts to obtain clearer images." In addition, hospital facilities can plan over time which AI models to adopt, based on the number of annual operations that are carried out on each organ or anatomical area.

The free online app can be navigated by starting either from the individual organs, or from anatomical areas such as the chest, neck or abdomen. Once the particular item has been selected, the app will report all the existing AI models, which have been tested on the image datasets that are available. The models can be sorted by dataset, from the most to the least effective. You can also select very particular organs, such as individual vertebrae or individual cardiac ventricles. Another very interesting feature is the option of sorting the models according to their ability to generate images of tumours and lesions, including those resulting from strokes and ischaemia.

Some of the models in the app are generalist, while others are specific to an organ or anatomical structure. As Andrea Moglia went on to explain: "The generalist AI models used in the medical field are trained on a huge and extremely varied set of images of the human body. They have recently proved in many cases to be just as effective as the specialist models, deliberately designed to generate images of a particular organ using one or a few datasets. They therefore represent a turning point for the sector."

Doctors and technicians have long used AI models to provide images of organs or lesions. Andrea Moglia added: "The technical term is segmentation, a process that allows you to delineate a particular structure in a 2D image, in order to produce a 3D reconstruction." In the medical field, it involves combining various photos taken from radiographs or CT scans, and indicating the particular organ or lesion with a coloured line. Using AI models makes this process faster, and avoids human error or bias.

Pietro Cerveri, Luca Mainardi and Matteo Leccardi, also from the Department of Electronics, Information and Bioengineering at the Politecnico di Milano, contributed to this work as well. The project was partly funded under the NRRP by the Future Artificial Intelligence Research foundation (FAIR).