Getting the most from data: how to organize heterogeneous data for effective AI analysis to investigate Flash and Minibeam radiotherapy techniques

The result of the observed FLASH and Minibeam effects is obtaining tumor efficacy similar to that of conventional radiotherapy while significantly reducing radiation-induced toxicity on normal tissues. The incomplete understanding of the underlying mechanisms, combined with the lack of comprehensive knowledge regarding the quantitative dependence of beam and irradiated tissue parameters on the extent of these effects, hinders their clinical translation. Therefore, further multidisciplinary studies, including beam dosimetric characterization, in vitro and in vivo radiobiology, simulation and data modeling analysis, are essential for effectively translating these promising techniques into clinical practice. In the PNRR-THE and INFN-MIRO projects, experiments on beam delivery, dosimetry, radiobiology, and modeling are being conducted, generating diverse data types (such as microscopy images, tabular, textual and raw data) and various parameters, making the study of radiobiological mechanisms a complex multivariate problem. While multiscale models have been developed to describe radiobiological responses, they require parameter optimization, which can be addressed with Machine Learning (ML) using experimental and simulation data. Artificial Intelligence techniques like clustering and neural networks can further identify patterns and hidden associations, with the added benefit of explainability methods to analyze the influence of beam and biological parameters on the radiobiological effects. To facilitate this multivariate analysis, the collected data of different origins and typologies need to be interconnected and adequately organized. We realized a modular database platform to store, manage, share and analyze the data collected within the two projects. The final aim is to have super-users that can access these data to perform correlation analysis through ML, simulations and data modeling techniques to comprehend the radiobiological mechanism. This centralized AI-based IT platform can play a crucial role in enhancing both the efficiency and speed of traditional preclinical studies and also improving data standardization and interoperability among various research institutions, undoubtedly accelerating clinical translation.