Radiomics: a great potential to impact the future landscape of digital healthcare
Following the three paper recommendations presented on Graylight Imaging LinkedIN profile over the last few weeks, we would like to focus on the concept of radiomics in more detail. At first we are going to introduce basic definitions and pipelines, then we would like to dive into the current challenges and opportunities of this promising field of research. We invite you to take part in the discussion about the potential risks and future prospects of radiomics.
The intense growth of modern technologies resulted in a constant increase of digitalized information being collected during a clinical routine. That opened new opportunities for analysis, but the amount and complexity of data required novel methods of automated processing. Radiomics is an emerging data-driven field of science focusing on the development and analysis of biomarkers extracted from medical images. The introduction of advanced machine-learning techniques led to the rapid development of radiomics research by providing tools for successful detection of patterns that were not available earlier to the radiology specialists through manual examination. Automated extraction of large amounts of quantifiable features may be used for disease diagnosis, survival prediction, the assessment of response to a treatment, and many more .
Radiomic features encode complementary biologic information from anatomical (e.g., MRI, CT) and functional (e.g., PET, fMRI) imaging. Features may be obtained by hand-crafted algorithms (figure 1) or through deep learning techniques which benefit from automated representation learning. Such hand-crafted and deep features may be split into the following groups based on their characteristics :
- shape-based features – descriptors of 2D or 3D binary segmentation masks
- first-order statistical features – describe the distribution of voxel intensities within the segmented image region, e.g., organ or tumor
- texture features – describe complex relationships between multiple voxels of the image
- other – features built upon transformed input images (e.g., by extraction of local binary patterns) or by novel custom algorithms
Figure 1: Liver Surface Nodularity Score as an example of hand-crafted feature for assessment of liver cirrhosis in CT image 
A typical radiomic pipeline consists of the following steps :
- image data acquisition and reconstruction
- segmentation of the region of interest
- feature extraction and selection
- data analysis
Figure 2: Example of radiomic pipeline for survival prediction 
Radiomic workflows vary according to a particular application or desired clinical outcome. The example of radiomic pipeline that uses structural MR data to estimate survival prediction has been presented in figure 2. Real-life multi-disciplinary radiomic pipelines incorporate radiomic features in many various ways: as data augmentation tool to feed a neural network with more training samples, to explain features produced by deep learning algorithms , or to combine deep learning and hand-crafted features in a radiomic-guided approach .
Radiomics has a great potential to impact the future landscape of digital healthcare:
- Hand-crafted radiomic features are well-defined and easier to interpret by a trained specialist. They reflect important clinical properties of anatomical structures by transforming original image information. Such features may be generated uniquely for the specific task by custom algorithms to obtain better and more reliable results.
- Though used mostly in oncology, there are many non-oncologic research areas where radiomic analysis proved effective, such as: research on Alzheimer’s disease , multiple sclerosis , and lung diseases . A typical radiomic pipeline relies on ground-truth segmentations to extract features from the masked scans. It is not always the case though – the preparation of high-quality ROI segmentations is a time-consuming and laborious task, and there are some specific applications for which the analysis of full 3D scans may have advantage over the standard approach with masked data.
- It is a non-invasive method for screening and clinical assessment of anatomical structures as a whole. This approach is especially useful for examination of tumors – due to their heterogeneous nature, biopsy may not always hit the change and provide reliable results. Cancerous lesions may grow in sensitive hard-to-access areas, so radiomic analysis may be used to assess the tissue and avoid invasive intervention. A great example of such approach are brain tumors: brain biopsies are strenuous for the patient with risk of possible complications. There are several features that may be used as robust indicators of brain lesion types  or for response assessment to the therapy . There are many constraints regarding such algorithms: consistent acquisition protocols must be precisely followed to ensure repeatable and reliable results. This is the case especially for MR studies since voxel intensities in this modality are not represented in standardized units.
- Radiomic-based applications may be used as a support tool for clinical decision-making in immuno- and radiotherapy . Therapy prognosis and survival prediction are some of the hardest tasks in clinical routine, burdened with great responsibility toward patient’s well-being. Radiomic approach is successfully used in clinical studies that require the assessment of response to the therapy .
- Once used for the treatment assessment, why not take a step further by selecting in advance the optimal treatment that would work best for a particular patient? That way we could tailor a specific therapy to fulfill needs of an individual and maximize the therapeutic effect . Such set of biomarkers is referred to as radiomic signature of a patient  , and may be used as a first step toward personalized medicine. Unfortunately, the developed signature is often limited to patients who are scanned with the same scanning parameters used to create the prediction model.
Besides undeniable highlights, radiomics faces many challenges that need to be addressed before the new approach is incorporated into day-to-day clinical practice. A radiomic pipeline is a complex system consisting of many processing steps, all of which require standardization and transparency. When these requirements are not met, errors tend to accumulate. Robust end-to-end evaluation procedures must be implemented along with test-retest approach in clinical environments to provide the final evidence about radiomics efficacy. The following limitations have already been identified  :
- Retrospective studies – one of the basic problems pertains to the historical aspects of radiomics development. There are various open-access datasets available for research, but many of them had been acquired during retrospective clinical studies. They did not follow standardized procedures or protocols, and many of them used some arbitrarily chosen cut-off values or parameters. A solution would be to plan prospective studies while having the radiomic angle in mind from the beginning.
- Heterogeneous cohorts – often precise information about tested cohorts is missing. We need to make sure that patients are diverse, so that prediction model is robust and generalizes well on new data.
- Insufficient standardization – the entire pipeline needs to be carefully standardized: from data acquisition, through pre- and postprocessing, to extraction of well-defined features and evaluation using commonly-used metrics.
- Limited amount of data – in medical field the amount of extracted features is often much higher than the available patient data, and extensive feature reduction/selection needs to be applied to avoid the curse of dimensionality.
- Interpretation of results – the lack of well-established prognostic and predictive references makes it impossible to evaluate results reliably and assess which features are truly relevant clinically. Possible mistakes may result from incorrect interpretation of the results (e.g., causation vs. correlation) or flawed analysis of data.
- Technical integration – seamless integration with hospital workflow is necessary, so that the radiomic pipeline is fully incorporated into daily routine of radiology specialists. This is crucial to collect important feedback and to not lose the clinical perspective.
One of the most critical issues is the standardization of every step of the processing. Different acquisition protocols used by different institutions result in different data obtained for the same patient. The scanning protocols have a great impact on generated data, especially inconsistent acquisition parameters of the scanners, such as: reconstruction kernel for CT, magnetic field settings and lack of standardized units for MR. Another story are time-dependent variables, such as: contrast dosage and timing – difference in timing between the injections of contrast agent affects voxel intensities and may result in generation of different images even for the same scanner. Other burdensome parameters have been presented in figure 3. To resolve that issue several methods have been proposed, such as: modeling the relationship and applying corrections accordingly , deep learning approach , post-reconstruction batch harmonization in order to harmonize radiomic feature sets by ComBat method .
Figure 3: Technical factors impacting different steps of radiomic pipeline 
To ensure repeatability and consistency of radiomic features the following initiatives have been introduced:
- Image Biomarker Standardisation Initiative (IBSI) to standardize mathematical formulas of feature definitions, processing steps, and resampling techniques : https://theibsi.github.io/
- Radiomics Quality Score (RQS) checklist: https://www.radiomics.world/rqs2
- Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement : https://www.tripod-statement.org/
Another approach is to create datasets tailored solely for verification of protocol consistency. There are publicly available datasets like the RIDER dataset to help to assess the impact of varying factors in radiomics. The public phantom dataset, designed for radiomics reproducibility tests on CT, helps to assess the impact of acquisition settings in order to eliminate non-robust radiomic features .
Radiomics – future perspectives
Radiomics is an interesting and promising field of research, but there are still many unanswered questions regarding technical and regulatory aspects. It needs to be incorporated into prospective studies to prove its value in a real clinical environment. The implementation of robust validation protocols will be the key to reproducibility, transparency and standardization of the entire processing pipelines. In my opinion, the hand-crafted radiomic features should be used along with the deep learning techniques to enrich current applications and provide us with new exciting solutions in the future.
We have gained experience in quantitative image analysis through development of our own deep learning-powered system for processing of DCE-MRI studies. This application will be described on Graylight Imaging LinkedIN profile in one of the next posts, and we will have another opportunity to share our insights regarding radiomics and quantitative image analysis.
 W. Rogers et al., ‘Radiomics: from qualitative to quantitative imaging’, BJR, vol. 93, no. 1108, p. 20190948, Apr. 2020, doi: 10.1259/bjr.20190948.
 J. J. M. van Griethuysen et al., ‘Computational Radiomics System to Decode the Radiographic Phenotype’, Cancer Res, vol. 77, no. 21, pp. e104–e107, Nov. 2017, doi: 10.1158/0008-5472.CAN-17-0339.
 A. D. Smith et al., ‘Liver Surface Nodularity Quantification from Routine CT Images as a Biomarker for Detection and Evaluation of Cirrhosis’, Radiology, vol. 280, no. 3, pp. 771–781, Sep. 2016, doi: 10.1148/radiol.2016151542.
 J. E. van Timmeren, D. Cester, S. Tanadini-Lang, H. Alkadhi, and B. Baessler, ‘Radiomics in medical imaging—“how-to” guide and critical reflection’, Insights into Imaging, vol. 11, no. 1, p. 91, Aug. 2020, doi: 10.1186/s13244-020-00887-2.
 F.-S. Ouyang et al., ‘Exploration and validation of radiomics signature as an independent prognostic biomarker in stage III-IVb nasopharyngeal carcinoma’, Oncotarget, vol. 8, no. 43, pp. 74869–74879, Aug. 2017, doi: 10.18632/oncotarget.20423.
 R. Paul et al., ‘Explaining Deep Features Using Radiologist-Defined Semantic Features and Traditional Quantitative Features’, Tomography, vol. 5, no. 1, pp. 192–200, Mar. 2019, doi: 10.18383/j.tom.2018.00034.
 H. Cho et al., ‘Radiomics-guided deep neural networks stratify lung adenocarcinoma prognosis from CT scans’, Commun Biol, vol. 4, no. 1, Art. no. 1, Nov. 2021, doi: 10.1038/s42003-021-02814-7.
 Q. Feng and Z. Ding, ‘MRI Radiomics Classification and Prediction in Alzheimer’s Disease and Mild Cognitive Impairment: A Review’, Curr Alzheimer Res, vol. 17, no. 3, pp. 297–309, 2020, doi: 10.2174/1567205017666200303105016.
 G. Pontillo et al., ‘A Combined Radiomics and Machine Learning Approach to Overcome the Clinicoradiologic Paradox in Multiple Sclerosis’, American Journal of Neuroradiology, vol. 42, no. 11, pp. 1927–1933, Nov. 2021, doi: 10.3174/ajnr.A7274.
 A.-N. Frix et al., ‘Radiomics in Lung Diseases Imaging: State-of-the-Art for Clinicians’, J Pers Med, vol. 11, no. 7, p. 602, Jun. 2021, doi: 10.3390/jpm11070602.
 Z. Yi, L. Long, Y. Zeng, and Z. Liu, ‘Current Advances and Challenges in Radiomics of Brain Tumors’, Front. Oncol., vol. 0, 2021, doi: 10.3389/fonc.2021.732196.
 A. Ibrahim et al., ‘Radiomics for precision medicine: Current challenges, future prospects, and the proposal of a new framework’, Methods, vol. 188, pp. 20–29, Apr. 2021, doi: 10.1016/j.ymeth.2020.05.022.
 B. Lou et al., ‘An image-based deep learning framework for individualising radiotherapy dose: a retrospective analysis of outcome prediction’, The Lancet Digital Health, vol. 1, no. 3, pp. e136–e147, Jul. 2019, doi: 10.1016/S2589-7500(19)30058-5.
 N. Papanikolaou, C. Matos, and D. M. Koh, ‘How to develop a meaningful radiomic signature for clinical use in oncologic patients’, Cancer Imaging, vol. 20, no. 1, p. 33, May 2020, doi: 10.1186/s40644-020-00311-4.
 A. Zwanenburg, ‘Radiomics in nuclear medicine: robustness, reproducibility, standardization, and how to avoid data analysis traps and replication crisis’, Eur J Nucl Med Mol Imaging, vol. 46, no. 13, pp. 2638–2655, Dec. 2019, doi: 10.1007/s00259-019-04391-8.
 J. Choe et al., ‘Deep Learning-based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or Masses’, Radiology, vol. 292, no. 2, pp. 365–373, Aug. 2019, doi: 10.1148/radiol.2019181960.
 F. Orlhac et al., ‘A Postreconstruction Harmonization Method for Multicenter Radiomic Studies in PET’, J Nucl Med, vol. 59, no. 8, pp. 1321–1328, Aug. 2018, doi: 10.2967/jnumed.117.199935.
 A. Zwanenburg, S. Leger, M. Vallières, and S. Löck, ‘Image biomarker standardisation initiative’, arXiv:1612.07003 [cs, eess], Dec. 2019, doi: 10.1148/radiol.2020191145.
 G. S. Collins, J. B. Reitsma, D. G. Altman, and K. G. M. Moons, ‘Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD): The TRIPOD Statement’, Ann Intern Med, vol. 162, no. 1, pp. 55–63, Jan. 2015, doi: 10.7326/M14-0697.
Contact us if you have any questions!
See the previous post: The Beauty of Seeing Beyond the Visible: Automating the DCE-MRI Analysis of Brain Tumors