Stepping stones towards a more quantitative future for oncology imaging

October 30, 2018 | Radiomics

A recent Red Journal article, by Olivier Morin from the University of California San Francisco and co-authors from the USA, France and The Netherlands, paints an enlightening picture of oncology’s progress towards fully harnessing the promise of quantitative imaging and radiomics.

The article highlights why high-tech quantitative analysis emerging from radiomics research is undoubtedly a part of future practice, but with a call-to-arms for those involved to take the proper steps to ensure we get there responsibly and in a way that will encourage adoption.

Radiomics, like genomics, is a big data discipline that relies on the accumulation of large amounts of information from past patients to build predictive models. The most robust models will be derived from large, high quality, well-curated datasets. More data means an improved ability to detect correlations between input image data and output outcome data. Input data can also include clinical, genomic and other data that may strengthen the models. Furthermore, to ensure predictions are widely applicable, models should be built using data from multiple sources, based on diverse sets of patients. It is essential that each contribution to the overall data pool adds rather than subtracts value (the garbage-in, garbage-out principle applies). Regarding imaging data, standardization of the imaging chain from acquisition to analysis to reporting is the key.

The Morin article highlights several efforts aimed establishing high quality data acquisition and analysis processes. For example, the Image Biomarker Standardization Initiative (IBSI) is active in consensus building within image processing, radiomic feature extraction and reporting. The TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) group issued the TRIPOD Statement as a guideline specifically designed for the reporting of studies developing or validating a multivariable prediction model. The National Cancer Institute’s Quantitative Imaging Network (QIN) provides recommendations for multi-site clinical trials and is evaluating standardized imaging approaches, analysis, display and clinical workflow.

The work of these groups and others is part of the drive towards improving the value and practicality of quantitative imaging biomarkers. The Quantitative Imaging Biomarkers Alliance (QIBA), has defined a QI biomarker as “an objective characteristic derived from an in vivo image measured on a ratio or interval scale as indicators of normal biological processes, pathogenic processes, or a response to a therapeutic intervention.” An insightful way to think about the efforts of these groups is that they aim to transform imaging devices into measurement devices via consistent, predictable data collection and analysis.

At HealthMyne, part of the internal product validation process is to test its image analysis tools against publicly available datasets and specifications. Furthermore, customers are encouraged to carry out their own tests. An example is use of the QIBA profile for CT Tumor Volume Change for Advanced Disease (CTV-AD), which lays out specifications for accuracy, repeatability and bias/linearity of volume measurements, using the various image sets made available. Such carefully prepared documents and datasets represent an extremely valuable service to the medical imaging community, spurred on by a common desire for quantitative imaging to break out of the research realm into mainstream clinical practice.

The Morin article points out in conclusion that a tangible first step could be to fully incorporate simple metrics such as tumor size and shape into routine clinical practice. Although radiologists routinely dictate tumor diameters as part of their report on a patient’s response to treatment, this is usually in aid of painting a descriptive picture of the situation for the referring physician. Rarely are these numbers placed in a digital format that can readily be used, for example, to automate the distinction of response vs progression vs stable disease via criteria such as the Response Evaluation Criteria in Solid Tumors (RECIST). Nor in that case can these values be mined by researchers in their ongoing quest for the most appropriate assessment of objective tumor response.

A step beyond tumor diameter is its volume. A change in volume is in principle a closer surrogate to the change in number of cancer cells harbored by a tumor. Indeed, one of QIBA’s main initiatives is to help legitimize the use of the volumetric response assessment via providing profiles that specify requirements for each step in the imaging chain. Accurate and consistent volume measurements at each time point before and after treatment requires tight quality control of each link in this chain.

One essential element of volume measurement and any other quantity pertaining to the pixels within a defined region is reliable tumor segmentation, itself a major ongoing area of research and development. For example, HealthMyne’s Rapid Precise Metrics (RPM)™ algorithms include segmentation based on a sophisticated pixel-by-pixel assessment of a tumor and its surroundings. RPM uses a combination of a user performing a quick “drag” of the cursor to indicate the approximate extents of a visible object such as a tumor, and a probabilistic method of classifying pixels as belonging to that object, or not. This represents a consistent, algorithmic, approach to determining the edge of an object, while still leveraging the input of an expert user.

Throughout the industry, major initiatives are also underway to apply deep learning techniques to segmentation. One important distinction between volume-based assessment and use of other more abstract radiomic features is that this simple metric is highly understandable. The leap of faith for a radiologist or oncologist in trusting a plot of volume change over time as an indicator of response to treatment is far smaller than putting trust in a model based on mathematically-derived features that can’t be visualized. Hence one of the challenges for researchers is to tie their potentially valuable models back to biological reality and thus encourage adoption.

With the combined efforts of the clinical, research and industrial communities we are heading towards a new era in medical imaging where “images are not just pictures, they are data” (Robert Gillies, 2016). However, let us give the final words to an important point from the Morin article:

“To fully harness the potential benefits of these immense sources of data and state-of-the-art techniques, but most of all, for a faster translation of quantitative imaging-based decision support into the clinical environment, it is crucial that all members of the medical imaging community act together to create worldwide consortiums, with the aim to improve practices, standardization, meaningfulness, usefulness, data sharing and clinical trial designs.”

Click here to read the full article.

Learn how the HealthMyne platform provides better metrics for tracking cancer lesions such as volume, density, mass, and more in this short video.

Intrigued? Let’s talk about what’s important to you.