Radiomics and Imaging: The Need for Standardisation

Encouraging recent clinical research in the field of radiomics has led to an increasing desire to translate radiomic image analysis into routine clinical practice. However, this process has to be preceded by an extensive knowledge of the robustness and reproducibility of the respective quantitative imaging features.

Unfortunately, an increasing number of studies has demonstrated so far, that these features are highly influenced by the technical settings used for image acquisition, reconstruction, and post-processing and that most of the features are not very robust when derived from images acquired with different technical settings.

As a consequence, the lack of standardisation in medical imaging represents the major hurdle for radiomics to be overcome in the future. Without achieving standardisation at the major points of the radiomics pipeline, a valid translation of radiomics to clinical practice will not be possible.

Thanks to a large (and still increasing) number of interesting radiomics studies over the past few years, we have seen many different fields where radiomics might add information regarding diagnosis, therapy monitoring or prognosis of individual patients, thus enabling imaging to proceed on its way towards precision medicine.

The largest field in imaging, where radiomics has been widely applied so far, is oncology. We all know that most tumours are not composed by a single tumour cell clone, but rather are a “mosaic” of genetically different subclones. These subclones might also change again during therapy due to mutations. These phenomena are subsumed under the header “intratumoural clonal heterogeneity.” What we currently do in radiology, is measuring sizes, at least when it comes to the standard criteria used in clinical studies such as the RECIST criteria. We all know the limitations of these criteria, and it is easily imaginable, that much information, which is in our images, gets lost while only measuring the size of a tumour. It has been shown, that radiomics might be able to detect the underlying intratumoural heterogeneity in our images and shows correlations to histopathology. Therefore, one wide field where radiomics might benefit the patient is the field of precision oncology, where the diagnostic and therapeutic management might be influenced by radiomics and where the non-invasive assessment of a tumour might become more important in the future.

Radiomics can add also prognostic information, which might benefit our patients. It is important to understand, however, that radiomics is not only applied in the field of oncology. There are various other fields, where radiomics is an interesting new aspect of imaging, such as in neuroimaging, musculoskeletal imaging or cardiovascular imaging.

I am very much engaged in the field of cardiovascular imaging and radiomics, where I am trying to improve the non-invasive diagnosis of myocardial inflammation, for example. We can show in some first proof-of-concept studies, that radiomics might add very relevant information about the (in this case inflammation-induced) inhomogeneity of the myocardial tissue, thus leading to potential novel biomarkers for a more precise diagnosis of active myocardial inflammation.

The special challenge of MRI is the relative nature of signal intensities depicted on standard MR images. In contrast to the more or less standardised Hounsfield Units in CT or the absolute SUV values in PET, these relative signal intensity values with their inherent variations impose a huge problem to the extraction of standardised, robust, and reproducible radiomic features.

We could demonstrate in a phantom study, that the robustness and reproducibility of radiomic features strongly depend on the MRI sequence used for image acquisition (eg FLAIR sequences delivering a much higher amount of robust features than T1w sequences) and the used reconstruction matrix size. Thus, the large amount of available MRI sequences for different scanners, field strengths and vendors imposes a huge challenge to feature standardisation.

Currently, it is nearly impossible to perform reliable radiomics studies on MRI datasets, since the test-retest robustness (which is one of the most important factors when it comes to clinical applicability) varies so much between the various acquisition settings. In my opinion, each radiomics study in MR should include an analysis of the robustness of the used features in the selected imaging setting in order to allow the reader to assess the generalisability and translational impact of the study.

AI with its subfields ML and DL already plays a role in “standard” radiomics analyses, since the large amount of quantitative data (“Big Data”) requires techniques from this field for statistical analysis. Usually, traditional statistical procedures are not capable to deal with the large amount of data, and ML techniques are required to reduce the feature set and select the most important features for a certain classification task.

In addition, ML and especially DL might be of special importance when it comes to the challenging step of achieving standardisation. Since the standardisation at the level of radiomic feature extraction appears to be challenging (at least for MRI), standardisation might be achieved through image harmonisation. First studies have been performed demonstrating the capability of DL techniques to achieve harmonisation of MR images acquired with different acquisition and reconstruction settings. This is a very interesting field, which has to be pursued in future research.

Finally, there is of course the possibility of applying ML/DL techniques directly to the acquired images, without the intermediate step of “handcrafted” radiomic feature extraction. In my opinion, however, this approach suffers from the same limitations as a traditional radiomics approach, and the lack of interpretability of the trained models hinders its translation to clinical practice.

For CT/PET-CT, the Image Biomarker Standardisation Initiative (IBSI) has been formed by Zwanenburg and colleagues, who provided valuable recommendations towards standardisation of the feature extraction process in radiomics. Unfortunately, there is no such initiative focused on MRI with its inherent challenges. That was actually the reason why my group started to focus on investigating the robustness of radiomics in MRI.

Currently, we and other groups are working intensively on this topic, potentially leading towards valuable approaches for standardisation. However, currently, there are no “ready-to-use” solutions available yet. Nevertheless, we are starting to better understand the influencing factors on robustness and reproducibility of radiomic features in MRI, and this is the first crucial step in the process of standardisation.

Quantitative MRI techniques are extremely interesting for radiomics research, since they do not suffer from the inherent limitations of MRI sequences displaying relative signal intensities. T1 or T2 relaxation times derived from T1/T2 mapping, for example, sometimes are called the “Hounsfield Units of MRI.” Thus, quantitative MRI parameters might be more robust when it comes to radiomics analyses than standard qualitative MRI images. However, quantitative MRI techniques also suffer from several unsolved limitations, which again have to do with lack of standardisation – for example, values are not comparable across scanners, field strengths or different sequences. We are currently investigating quantitative imaging techniques regarding robustness and their potential to deliver more reproducible radiomic features.

Yes, this is possible. If we do not manage to solve the standardisation problem, there will be too much variation in the features, which will hinder a valid translation into the real-life clinical setting.

In clinical studies, it might be feasible to adhere to constant acquisition parameters, but this will not be possible in the real-life setting. However, we are facing similar problems with CT and PET, so, I am not sure if this is true only for MRI and radiomics.

Well, radiomics is basically a bunch of numbers. The interpretation strongly depends on the individual disease and setting. Sometimes, it is possible that only one or two combined features allow classification between healthy and diseased, while most of the times, there is a set of several (between 5 and 20, sometimes much more) features needed for classification.

In those cases, an ML algorithm is usually trained and validated in order to perform the classification task. It appears to be feasible that in the future, once the current challenges have been addressed and the standardisation problem has been solved, these trained algorithms will be integrated into the diagnostic decision making pipeline such as in a CAD system, and the trained algorithm then classifies images into predefined categories.

To me, it does not appear feasible that the radiologist himself has to calculate a bunch of numbers, makes some advanced statistics and then combines the numbers into a meaningful diagnosis. Techniques like radiomics will only be applicable in routine clinical practice if fully integrated into the clinical workflow and automated as much as possible. This still is a long way to go.

Firstly, I would be very happy to see fewer radiomics studies performed and published with poor methodology and instead adhering to the quality guidelines, which have been published in recent times. Radiomics studies with ML methodology and without at least a small separate testing dataset should not be published anymore. Of course, most of us started with small studies without any validation or testing datasets, but since there is increasing awareness of the limitations including lack of generalisability of such studies, we should adhere to stricter standards.

Secondly, I would like to see more researchers and clinicians accepting the current limitations of radiomics and the concomitant need to address the issue of standardisation. Hype does not contribute to a valid clinical translation, and if radiomics is translated to clinical practice too early without addressing the limitations, the technique will be dead before we had the chance to solve all the problems.

Finally, I would like to see more studies investigating the influence of various image acquisition, reconstruction and post-processing settings on the robustness of radiomic features and more efforts to address the standardisation problem. This might also be something, where the different vendors might contribute (but I am rather skeptical if they want to standardise their protocols and sequences). Therefore, I guess we will have to find other solutions for harmonisation and standardisation of the images underlying radiomics analyses.