By Rebecca Montz, EdD, MBA, CNMT, PET, RT(N)(CT), NMTCB RS
Vol. 24 No. 6 P. 14
Using Radiomics to Go Beyond Human Sight
Radiomics in medical imaging is a groundbreaking quantitative approach that extracts features from acquired images using advanced mathematical analysis. The quantifiable features extracted from medical imaging data are based on pixel intensity, pixel arrangement, pixel color, and texture. The use of radiomics allows the enhancement of existing data by implementing data-characterization algorithms. The goal of radiomics is to maximize imaging data to detect beyond what the human eye is capable of detecting.
Gary Cohen, MD, chair of the department of diagnostic imaging at Temple University Health System in Philadelphia, explains that radiomics is the use of big data to collate information from imaging studies such as CT and correlate it with pathologic correlation, molecular data, and other parameters to identify patterns of disease not seen by ordinary imaging and analyses. He says scientists and radiologists have the potential to build a large catalog of imaging patterns and recognition to predict disease growth, spread, and tissue type early in the cellular process, which will hopefully result in prognostic significance and enable early treatments as well as more focused and specific treatments based on radiomic determination of tissue types.
Currently, there is no FDA-cleared radiomics algorithm for clinical use in the United States; however, there is published research utilizing radiomics in medical imaging that shows its potential to enhance clinical decision making and potentially improve patient care. One area that has shown promise is oncology, due to radiomics’ ability to determine tumoral patterns and characteristics possibly not noticed in patient images. Although published studies using radiomics have demonstrated positive outcomes, there are still challenges caused by various technical factors influencing the extracted radiomic features. Cohen is optimistic that radiomics is perhaps the next big horizon for radiology and imaging to tackle.
The Lay of the Land
Currently, radiology has embraced many facets of AI to result in faster diagnosis of diseases that require prompt recognition and treatment. Through AI processes, radiologists are now able to seamlessly recognize critical results as diseases are brought to the attention of radiologists quickly, without work disruption, by identifying patients with positive findings needing immediate attention, Cohen explains. Traditionally, radiologists’ worklists are ordered chronologically, but they can now be triaged to identify findings requiring immediate attention. Patients with acute pulmonary embolus and right heart strain, intracranial hemorrhage, perforated bowel, rib fractures, dissections, and stroke, among other conditions are being brought to the prompt attention of radiologists, he adds.
Both radiomics and AI applications have contributed to technological advances in medical imaging; however, there are differences between the two. Marwan Sati, PhD, distinguished engineer for Merge by Merative, a data, analytics, and technology partner for the health industry, explains that radiomics focuses on the extraction and analysis of quantitative features from medical images (eg, tumor size and texture), whereas most AI applications help interpret images (eg, lesion detection and classification). Sati says both technologies offer opportunities to help radiologists improve their workflow and performance efficiency.
Cohen adds that radiomics will further AI in the health care fields, which is still in its early stages, so that extremely large data sets and algorithms will enable clinicians to identify tumor characteristics and other disease characteristics that could portend diseases before they are recognizable without these processes. “By correlating these subtle patterns of imaging with pathologic outcomes, we hope to impact overall survival with early diagnosis, interventions, and treatments,” he says.
Radiomic data are commonly derived as features from images, using mathematical principles and advanced image processing algorithms. Although any CT or imaging modality can be processed with radiomics software, radiomics output varies by technical variations in acquisition and reconstruction parameters. According to Jia Wu, PhD, of the department of imaging physics at the University of Texas MD Anderson Cancer Center, there are several key steps to obtain radiomic data, including study cohort building, DICOM data downloading from PACS systems, image data harmonization, region of interest contouring, radiomics feature extraction, and machine learning model building and testing. He emphasizes that it is important to maintain some level of consistency between the parameters to ensure stable performance of radiomics for a given task, such as lesion characterization or prediction of treatment response.
There are several research algorithms, both commercial and open access, that users can deploy on their hardware. Sandeep Bodduluri, PhD, an assistant professor of medicine at the University of Alabama at Birmingham, explains that a radiomics feature extraction workflow usually consists of input medical images fed through an image segmentation algorithm to focus on a specific region of interest within the provided images. Feature extraction involves collecting various measures related to image intensity, texture, and shape of the organ of interest. The large feature pools are then processed through a machine learning/ AI algorithm for disease detection and prognostication tasks. Mannudeep Kalra, MD, a professor of radiology at Harvard Medical School, adds that algorithms vary in their segmentation capabilities; some require manual drawing of regions of interest (ROI), while others have semiautomatic or automatic ROI capabilities. Most radiomics provide similar data with several hundred radiomic features. Some radiomics can only process single energy CT images, while others can process both single-energy and dual-source, dual-energy CT images, Kalra says.
As previously mentioned, there is no FDA-cleared radiomics software for clinical use in the United States. However, in research settings, investigators have explored a variety of uses for radiomics in oncologic and nononcologic settings. Kalra says the most frequent use pertains to prediction of treatment response and prognosis, followed by prediction of genetic mutations associated with certain cancer types—such as lung cancer and colorectal cancer—and lesion characterization for differentiating benign and malignant etiology. Wu agrees that CT radiomics has been broadly used in bridging clinical gaps in the oncology field and has shown promising results in improving screening, diagnoses, prognosis, and treatment response access. CT radiomics has been correlated with pathological and genomic factors, which is generally referred to as radiogenomics analysis, such as using radiomics to predict lung cancer subtypes.
Radiomics research has shown promise in specialties outside of oncology as well, including lung disease. Bodduluri and colleagues’ recently published study, “Radiomics for Improved Detection of Chronic Obstructive Pulmonary Disease (COPD) in Low-Dose and Standard-Dose Chest CT Scans,” demonstrated high accuracy using both standard-dose and low-dose CT with the advantage of radiomics technology. Bodduluri explains that the study investigated the significance of radiomics features derived from inspiratory CT scans for the detection of COPD. The features represented lung shape, texture, and density, and they outperformed prior deep learning models with higher accuracy for detection of COPD.
Another recently published study in Radiology, a journal of the RSNA, “A Coronary CT Angiography Radiomics Model to Identify Vulnerable Plaque and Predict Cardiovascular Events,” utilized radiomics technology to predict future cardiac events. The study included researchers from China who developed a radiomics model that used information from coronary CT angiography images to assess plaque vulnerability. The model enabled the detection of vulnerable plaques associated with an increased risk for major adverse cardiac events, such as heart attacks.
Research has shown that the use of radiomics technology has the capability to help define, standardize, and foster big datasets for clinicians to connect patients with specific health care profiles all over the world. Bodduluri agrees that the field of radiomics is evolving, and although several new publications for a variety of applications are drawing attention to the utility of radiomics, challenges remain.
The last two decades of multimodality research into radiomics has highlighted many strengths and opportunities in both oncologic and nononcologic arenas, but the lack of a clear solution is a significant hindrance for mainstream applications of radiomics outside of research. Kalra says although integrated deep learning features within some radiomics can help with rapid, single-click analysis and result display, radiomics are still too complex for most clinical users. Most radiology interpretations, including for CT, are qualitative and subjective in nature, and there is no clear answer on how thousands of radiomics features and values can find their way into individual radiology reports, Kalra says. Other challenges are associated with variations in radiomics performance related to variations in scanning and/or patient factors.
Wu adds that building robust radiomics models with high reproducibility and consistent performance in CT data from different centers or vendors and interpreting radiomics models also contribute to the many challenges. “Logistically speaking, deploying a radiomics model into clinical flow is not easy and will require specific systems and support, including quality control, data preprocessing, model implementation, and interpretation,” Wu says.
According to Bodduluri, radiomics features from CT, as with any other image-based features, are sensitive to image acquisition parameters, such as the volume of inspiration, radiation dose settings, scanner model settings, and image noise. He also mentions that there has been a significant push toward validating the robustness of radiomics features across various acquisition and dose settings. Juan Carlos Ramirez-Giraldo, PhD, director of CT research and development collaborations at Siemens Healthineers North America, says lack of standardization and universal models are key reasons why radiomics has not advanced to clinical practice. He says universal guidelines, robust standardization, and ease of use will contribute to radiomics moving into future clinical use.
On the Horizon
Although technology cannot replace a clinician’s expertise or clinical judgment, radiomics can provide additional tools to improve the standard of care for all patients. Radiomics has the potential to improve workflows and accuracy of patient diagnoses and may reduce the amount of unnecessary invasive procedures. The use of radiomics technology is easily integrated into the workflow of CT imaging without administering additional radiation exposure or cost to the patient. Radiomics can aid radiologists in discovering unknown patterns or specific information overlooked by human eyes.
Kalra remains hopeful that technological advances such as generative AI and photon-counting detector CT will help raise the stature and potential of radiomics for clinical use. He says both limited and large dataset studies have provided compelling evidence for potential applications of radiomics on CT data, particularly for oncologic applications. However, lack of generalizability and reproducibility from a myriad of technical and patient factors, apart from their complexity, explain the lack of clinical adoption of radiomics. Despite the challenges, Kalra hopes people will soon be discussing how and when radiomics can become a clinical tool for noninvasive lesion characterization, predicting disease prognosis, and treatment response at the time of initial detection. Ramirez-Giraldo says the abundance of literature demonstrates enthusiasm around radiomics, especially with interest in oncology, and is hopeful that radiomics will transition from being an exciting research area to generating impact in clinical practice.— Rebecca Montz, EdD, MBA, CNMT, PET, RT(N)(CT), NMTCB RS, has worked at the Mayo Clinic Jacksonville and University of Texas MD Anderson Cancer Center in Houston as a nuclear medicine and PET technologist.