In radiomics, large numbers of quantitative features from images (e.g. CT or MRI scans) are extracted and analyzed. These image features cannot be identified by the human eye, i.e. there is much more information in an image than intuitively appreciated. These radiomic features can be used for many applications in medicine, such as prognostication of individual patients, as a predictive test, for response assessment, for automatic contouring of anatomical structures including tumors etc.
Radiomics and deep learning are not the same, but have overlapping characteristics, of which the ability of deep learning to select features independent of human interaction is one of the most striking.
A typical radiomics pipeline involves the acquisition of images, segmentation of a region of interest (ROI), e.g. a tumor, feature extraction, statistics and predictive modelling and finally validation.
Radiomics is shown to give information that relates to biological characteristics of the tumor such as hypoxia, PD-L1 expression and CD8 T-cell infiltration. The prediction of these features was shown to correlate with the survival of patients.
In line with the biological correlate of radiomics, response evaluation in situations where RECIST criteria perform modest such as in case of immune therapy, radiomics helps in classifying patients and correlates with survival.
Radiomics has therefore a role to play the clinical cycle where at several points decisions have to be made, with or without patient involvement, e.g. prognostication of patients to allocate the optimal treatment, segmentation of organs at risk and tumors, response assessment etc. Moreover, as radiomics relies on standard images, including cone-beam CT scans that are made on a daily base in standard practice, the changes of the image over time in an individual patient will increase the predictive accuracy of radiomics.
Even though radiomics has shown to have big potential, the majority of publications are of inferior quality, being in over 90 % based on retrospective, i.e. highly biased, data sets, with again over 90 % single-center and mostly not externally (ideally at least twice) validated. Mostly, the AUC of the model is only 0.55-0.60, with significant heterogeneity. The latter may also partly be due to differences in scanners and image protocols.
Nevertheless, although radiomics is a novel scientific field, already at present, it has shown its big potential impact for it allows to give even standard non-contrast-enhanced Imaging insight in biological mechanisms and helps in prognostication and response prediction together with other known variables. The major caveat in a lot of radiomic literature is the lack of multiple external validation and reproducibility.
The exciting new data coming from radiomics will be part of the ongoing move towards AI-driven radiation oncology.
References
Lambin P, et al. Radiomics: extracting more information from medical images using advanced feature analysis.
Eur J Cancer. 2012 Mar;48(4):441-6
Coates JTT, et al. Radiomic and radiogenomic modeling for radiotherapy: strategies, pitfalls, and challenges.
J Med Imaging (Bellingham). 2021 May;8(3):031902.
Grossmann P, et al. Defining the biological basis of radiomic phenotypes in lung cancer.
Elife. 2017 Jul 21;6:e23421.
Aerts HJ, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach.
Nat Commun. 2014 Jun 3;5:4006
Sun R, et al. A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study.
Lancet Oncol. 2018 Sep;19(9):1180-1191.
Dercle L, et al. Reinventing radiation therapy with machine learning and imaging bio-markers (radiomics): State-of-the-art, challenges and perspectives.
Methods. 2021 Apr;188:44-60.
Trebeschi S, et al. Predicting response to cancer immunotherapy using noninvasive radiomic biomarkers.
Ann Oncol. 2019 Jun 1;30(6):998-1004.
Sanduleanu S, et al. Non-invasive imaging prediction of tumor hypoxia: A novel developed and externally validated CT and FDG-PET-based radiomic signatures.
Radiother Oncol. 2020 Dec;153:97-105.
Hope A, et al. Artificial Intelligence Applications to Improve the Treatment of Locally Advanced Non-Small Cell Lung Cancers.
Cancers (Basel). 2021 May 14;13(10):2382