Radiomics: 3 Challenges That Must Be Solved Before Widespread Adoption Occurs

Radiomics: 3 Challenges That Must Be Solved Before Widespread Adoption Occurs

By Robert Holmes, Ph.D., Chief Science Officer, HealthMyne

Radiomics is a field of research in which image analysis is used to extract large number of quantitative features from medical images, such as MR, CT, and PET scans. While there are many potential applications of radiomics across the healthcare spectrum, oncology continues to be one of the main areas of study.

Radiomic features themselves range from simple tumor and lesion shape measurements that are standard practice for radiologists, such as lengths, through more detailed shape measurements that do not form part of the radiologists workflow, such as volumes and areas. Simple statistical measures are also made along with higher-order measures that describe the texture of the lesion, its fractal properties, or that focus on specific regions, e.g. the peritumoral.

Although more than 2,000 radiomic features can be calculated through AI-enabled technology for a single lesion, it remains challenging for researchers to determine which of these measures are useful. That has been the difficult task confronting radiomics researchers over the last decade, who have sought to map these features—alone or in combination with one another—to useful outcome. Examples of such models include predicting progression-free survival, the probability of response to immunotherapy, or the presence of a particular genetic biomarker.

How researchers use radiomics in practice
In radiomics, the analytics pipeline is similar to that in any other big data project: one of visualize—reduce—select—validate. Typically, a researcher makes an initial visual “sanity check” to identify any obvious errors, outliers, and patterns that will inform subsequent analysis. Next, dimensional reduction is used to identify and remove redundant data elements (highly correlated pairs of variables, for example), to estimate the inherent “true” dimension of the problem, and to project the data onto this smaller subspace.

Selection involves applying a range of statistical-learning and machine-learning algorithms and scoring their relative performance to identify the most successful model for this particular study. The model is then subject to rigorous validation on a “holdout” set of previously unseen data; only if it passes this step is a model deemed suitable for deployment.

Despite the promise of radiomics to enable researchers to uncover previously unobtainable insights into the biology of tumors and lesions, it is important to keep in mind that the field is subject to some significant restrictions:

Highly correlated features: Many radiomic metrics are strongly correlated with one another, and the inclusion of all of them does not add significant information to an analysis. In addition, many statistical techniques make assumptions about the independence of model inputs, so the researcher must be careful to either remove the correlated features or be careful in their selection of methodology.

Low volume of data: It is not unusual to see papers attempt to build models from records on a hundred subjects or fewer. When each of these records may have several thousand radiomic attributes attached to it, it is easy to fall victim to over-fitting, in which the model memorizes the data rather than generalizes from it. Deep learning models are notoriously greedy with regards to data and so—when using such techniques—particular care must be used whenever the subject count is measured in the hundreds and not the thousands.

Generalizability: The majority of radiomic features are innately heterogenous. Unlike a biological analysis in which a sample is obtained by simple extraction or excretion, images are obtained from complex equipment that are subject to variability from vendors, the radiologists themselves, and the machine settings for individual image acquisition. For researchers, the solution to this challenge is three-fold. First, is an increased focus on multi-site multi-machine studies, likely requiring collaboration. Second, is an adoption of the standards that are proposed being developed by bodies such as the Quantitative Imaging Biomarkers Alliance (QIBA) and the Imaging Biomarker standards Initiative (IBSI). Third, is an increased reliance on automation and semi-automation when identifying and segmenting lesions in the image.

In summary, the field of radiomics and its application to oncology is at a turning point. Ten years of literature have already given us many proof points of the potential of these features to inform oncology research. It is now time for the field to mature, for it to jump the “translational gaps” that will see it fully adopted as a medical research tool and subsequently as clinical decision-making tool. But this can only be done by robust and consistent responses to the identified challenges.

About the author:

Dr. Holmes has over 25 years’ experience in the successful application of machine learning and advanced analytics to a range of commercial and industrial problems. Since 2004, he has been working exclusively in the healthcare sector, where he has delivered solutions to pharmaceutical clients, healthcare delivery systems, and health plans.