By Hugo Aerts, Ph.D.
Director, Computational Imaging and Bioinformatics Laboratory (CIBL)
Associate Professor, Harvard University
In the past year, the use of Artificial Intelligence (AI) in radiology, also called “radiomics,” has been getting a lot of attention, mainly because of the progress Deep Learning (DL) has made from a sub-human performance to performance that is similar, or in some cases superior, to that of humans. This progress is propelling the field of radiomics forward at a rapid pace. As AI and DL continue to advance and out-perform humans at some tasks, and as we discover more about the complementary relationship between radiomics and other diagnostic approaches, such as genomics, we can expect the use of radiomics in clinical practice to grow dramatically. Yet, despite this progress, concerns about the future for radiologists are unfounded — their skills and insights will be needed for a long time to come. There are challenges related to radiomics that we must overcome, and if we look back over the last ten years, we find we can learn from other domains rather than repeat the same mistakes.
Radiomics presents many advantages over other diagnostic methods. Radiological testing can provide a 3D picture of a solid tumor, and is non-invasive, unlike traditional biopsies, so it can be performed multiple times without side effects. On the other hand, radiomics provides a macro view of the cancer, and therefore more qualitative, rather than quantitative, insight into the tumor. The acquisition protocols are also very heterogeneous, which makes comparisons between patients, or even within the same patient in time, difficult. The goal therefore is to determine the specific strengths of radiomics, and how those can combine with other methods and data, such as that from molecular assays, to improve human ability to detect, characterize, and monitor cancer.
Through the application of AI, features can be extracted from images that provide much more information than the human eye could discern. Imaging, for example, can capture macro differences between tumors, such as size, shape, and surface features (smooth versus rough and infiltrating cancers.) If these phenotypic features can be linked to specific mutations, for example, the information could be used to determine treatment or predict outcomes.
In one analysis my colleagues and I performed, we analyzed CT imaging of almost 1000 patients with lung or head and neck cancer, and we developed and validated a prognostic radiomic signature quantifying intra-tumor heterogeneity. This radiomic signature outperformed visual analysis on volume measurement and was complementary to the TNM Classification of Malignant Tumors (TNM) staging on all validation datasets. Further, the Imaging-Genomics analysis showed a strong correlation between radiomics and genomics data. In another analysis using an integrated set of data from over 750 lung cancer patients with somatic mutation testing and engineered CT image analytics, we looked at whether phenotypic features of a lung cancer tumor can predict KRAS/EGFR mutational status. We selected 26 features a priori, and found 16 features associated with EGFR mutations and 10 associated with KRAS mutations. EGFR-driven tumors had higher radiographic heterogeneity and smaller volume, presenting overall lower density. KRAS-driven tumors, on the other hand, were more likely to be homogeneous with a similar size compared to non-KRAS mutated tumors. Our analysis indicates that somatic mutations drive distinct radiographic phenotypes that can be predicted by radiomics.
These results are clearly encouraging and exciting, but within the domain of medical imaging, we need to address some challenges if we want radiomics to progress to the next level of scientific and clinical contribution to the fields of cancer research and care. Imaging algorithms must have high performance and stability, and very large datasets are required for validation. One problem we have is that many radiomic studies are currently very raw; the study and resulting data have not been designed in a rigorous manner. There is no standardization in terms of data and methods, and as a result, data sharing is not common. Also, many imaging researchers don’t understand the necessity of involving data scientists in the design of studies. The most amazing thing is that if you talk with genomic experts and bioinformaticists, this is exactly what happened in these fields 10 years ago — we are repeating the same mistakes! Yet it does not have to be the case.
In order to learn what we can from these other domains, I see three important points for our path forward. First, the imaging community must start sharing data and methods, as the genomics community does. Journals should demand that the data and code are made public, to ensure the ability to reanalyze data for new purposes, and to support reproducibility or integration with other work. Second, there need to be better initiatives towards standardization — QIN is a step in that direction, but a more focused effort is required, specifically around AI algorithm definitions, and data standards and sharing. Finally, we need to involve more data scientists in the radiology domain, and imaging researchers need more in-depth training in statistics and experimental design. There has been a large increase in studies with serious experimental design flaws and this can weaken the perception of the field (e.g., small datasets, lack of independent validation, flawed statistical analyses.) We need a strong foundation for the future, so we must ensure experimental design is rigorous and we that we are drawing correct conclusions.
We know that AI will have a huge influence on imaging and will change the profession. Some radiologists express concern that they will become obsolete, and imagine they will be replaced by smart algorithms and Deep Learning, but they needn’t be worried. Right now, despite the enormous promise of AI in imaging, there is a lot of hype that ignores the challenges I’ve mentioned. It will take time before large chunks of radiology processes are actually automated. This means that radiologists will be in the driver’s seat for a long time to come. If we act now as a community to drive the progress of radiomics in the right direction as a computational science, AI and DL will only enhance and solidify the role of radiologists as essential partners in clinical practice and cancer research.
Watch Dr. Aerts’ Nov. 8, 2017 presentation to the NCI CBIIT Speaker Series.