The next_AIM experiment aims to address the following specific challenges related to methodological aspects of the application of AI in Medicine (AIM):
challenge 1) – how to manage limited datasets with AI techniques (no-so-big dataset);
challenge 2) – how to make solutions provided by AI models understandable by humans (explainable techniques for AIM).
The following methodological approaches will be implemented to address the two main challenges of the project. Challenge I). Reliability and reproducibility of the results are mandatory for medical applications based on AI. Training AI models with limited annotated data samples poses specific challenges on the robustness and generalization ability of AI models. Specific guidelines should be defined regarding the definition of efficient training algorithms and rigorous cross-validation protocols either to enable the use of AI techniques in case of limited data availability for a specific study, or to discard the possibility of using them. Challenge II). The scientific community is currently actively working on a variety of methods for opening the AI “black boxes” to explain why an AI-based system has provided a specific response on a given input. This aspect is extremely important in the field of medical applications, and deserves specific technical developments, implementation and validation, in collaboration with clinical experts. Both challenges will be addressed first focusing on theoretical approaches and conceptual experiments, then experimenting on real data samples (public and private collections and their combinations).
