The review article “Artificial intelligence in cancer research: learning at different levels of data granularity” by Davide Cirillo and Iker Núñez-Carpintero, from the Computational Biology group led by Alfonso Valencia, is featured on the cover of the April issue of the Molecular Oncology journal. In this review article, BSC researchers survey the challenges and limitations of the state-of-the-art AI approaches for cancer research.
Artificial Intelligence (AI) is being used in a wide range of applications that aim at improving cancer diagnosis, prognosis and therapy. As an example, a number of radiological devices using AI have been recently approved by FDA for their medical application in oncology. Despite the high performances and great potential for the future of cancer medicine, AI systems are “data-hungry”, meaning that they depend heavily on large amounts of data for training. As a consequence, the size of the datasets that are needed to train AI models represents one of the major limitations in this area.
It is often very difficult to obtain datasets that are large enough to train complex models in many areas of cancer research. For instance, less common pediatric tumors, which affect a small number of children compared to the general population, are characterized by small sized datasets. However, the rarity of a condition is not always the reason behind the scarcity of large datasets. For instance, the more we disaggregate the data by cancer subtypes, or just demographic subgroups (age, race, gender), the more we reduce the size of the datasets that can be used.
These differences in the granularity of cancer data are challenging the application of AI in oncology. Indeed, the development of AI solutions that allow learning from small datasets is a much needed endeavor, especially for the realization of personalized medicine approaches based on AI. In this review article, BSC researchers address this issue by surveying challenges and limitations of the state-of-the-art AI approaches for cancer research. The authors discuss several techniques that enable the application of AI to small datasets, including transfer learning, meta-learning, semi-supervised learning and many others. Moreover, the authors propose a perspective view on the benefit of implementing synergistic solutions among different AI techniques, with special emphasis on synthetic data generation.
Article: Artificial intelligence in cancer research: learning at different levels of data granularity
https://doi.org/10.1002/1878-0261.12920
