More Than Meets the Eye
By Dan Harvey
Radiology Today
Vol. 20 No. 11 P. 12
Radiomics finds hidden features in imaging data.
The concept of radiomics moves beyond traditional visual interpretation. It involves stages that include image acquisition and subsequent reconstruction, segmentation, extraction, and qualification—all of which lead into analysis and eventual treatment model creation. This leads to better prediction of prognosis and therapeutic response, which, in turn, improves a patient-personalized treatment regimen. The concept has elicited great interest among physicians, particularly those treating cancer patients.
“Radiomics is computerized feature analysis of radiographic scans to capture quantitative phenotypic attributes of the tumor,” says Mohammadhadi Khorrami, MS, a PhD candidate from the department of biomedical engineering at Case Western Reserve University School of Engineering in Cleveland. “Most treatment decisions are based on clinical parameters such as cancer stage, genetic examination, age, gender, and, in some cases, ‘smoking status.’ These biomarkers are imperfect.”
But, he adds, radiomics can better predict beneficial response to treatment. Khorrami explains that radiomics-based analysis can provide more detailed characterization than possible by standard image viewing. His research into this area demonstrates effective usage for a diagnostic problem—distinguishing adenocarcinomas from granulomas—and prognostic problems, such as predicting response to different types of therapies for lung, breast, and prostate cancers.
Anant Madabhushi, PhD, F. Alex Nason Professor II of Biomedical Engineering and director of the Center for Computational Imaging and Personalized Diagnostics (CCIPD) at Case Western Reserve University, has been involved in radiomics-related research for nearly two decades. He defines radiomics as the high-throughput extraction of quantitative information from routine radiological images, such as X-rays, CT, MRI, and PET, for defining the textural and morphological characteristics of a given disease.
“Radiomics comprehensively and quantitatively characterizes, pixelwise, tumor characteristics on imaging,” he says. This can happen in several ways:
• shape features that measure how regularly or irregularly tumor boundaries change based on their 3D topology;
• semantic or qualitative features, which include radiologist-derived assessments of the tumor, including spiculations and the size of the tumor along several axes; and
• intratumoral heterogeneity measures, including gray-level features that investigate pixel level differences in the texture of the tumor in order to characterize its heterogeneity.
A Closer Look at Lung Cancer
Khorrami engaged in a study that was recently published in the RSNA journal Radiology: Artificial Intelligence; Madabhushi is the senior author. Khorrami describes what the team of researchers hoped to achieve.
“We wanted to see if patients with non–small cell lung cancer (NSCLC) treated with chemotherapy will respond before starting therapy by just looking at CT images, and we wanted to see if quantitative analysis of CT images can predict overall survival and time to progression in these patients,” he says. Access to CT images before therapy initiation appeared more predictive of response to chemotherapy.
Khorrami provides a glimpse of what was revealed. “Texture of tumor between responder and nonresponder are different, and this shows the biological difference inside the lesion,” he says. The researchers’ purpose was to identify how well radiomics texture features, both within and outside the nodule, predict the following:
• time to progression (TTP);
• overall survival (OS); and
• response to chemotherapy in patients with NSCLC.
The retrospective study involved 125 patients who had been treated with pemetrexed-based platinum doublet chemotherapy at Cleveland Clinic. Researchers randomly divided patients into two sets: a training set that included 53 patients with NSCLC and a validation set that included 72 patients. To predict chemotherapy response, the researchers used a machine learning classifier trained with radiomic texture features extracted from intra- and peritumoral regions of non–contrast-enhanced CT images. They used the Cox regression model, which generated the radiomic risk-score signature, for determining least absolute shrinkage and selection operator. In addition, they evaluated the association of radiomic signature with TTP and OS.
The researchers reported that “[a] combination of radiomic features in conjunction with a quadratic discriminant analysis classifier yielded a mean maximum area under the receiver operating characteristic curve [area under the curve, or AUC] of 0.82 ± 0.09 (standard deviation) in the training set and a corresponding AUC of 0.77 in the independent testing set.” They also indicated that the radiomics signature significantly associated itself with TTP (hazard ratio [HR], 2.8; 95% confidence interval [CI]: 1.95, 4.00; p<0.0001) and OS (HR, 2.35; 95% CI: 1.41, 3.94; p=0.0011).
Perhaps most significantly, the decision curve analysis revealed that—in terms of clinical viability—the radiomics signature had “a higher overall net benefit in prediction of high-risk patients to receive treatment than the clinicopathologic measurements.” Researchers concluded that radiomic texture features extracted from within and around the nodule on baseline CT scans are predictive of chemotherapy response and associated with TTP and OS for patients with NSCLC.
“[F]or this specific study, we showed that those patients who have more heterogeneous intratumoral and peritumoral texture will not benefit from chemotherapy,” Khorrasami says.
The study was unique in that, “for the first time, we showed that not only intratumoral features are important but also a small region around the tumor [peritumoral] has some clues that can show which patient will benefit from therapy,” Khorrami explains. “We also showed that those features that can predict response to chemotherapy are also associated with overall survival and TTP.”
The study’s conclusions lead to the inevitable question: Will it ultimately enable clinicians to do—or at least realize—anything they couldn’t do before? Khorrami’s response is emphatic.
“Yes, it will help clinicians know if a patient will or will not respond to therapy,” he says. The ability to determine, from a baseline scan, who is going to respond to therapy would be a game changer. “If we determine which patient won’t respond to a certain therapy, [physicians] can look at alternative therapies or combination therapy,” he adds.
Knowing who will respond to therapy can not only improve patient care but also help reduce costs. “We need tools where, in the noninvasive way, we can figure out who will respond to therapy, and this is both from the patient and economic perspective,” Khorrami says. Chemotherapy drugs cost approximately $30,000 per year, per patient.
Subvisual Clues
Madabhushi was the principal investigator in another significant radiomic-related study, published in a 2019 issue of Radiology. Participating colleague Niha Beig, a fourth-year PhD candidate in the biomedical engineering department at Case Western Reserve University, describes what the research team was looking for in their study, which was conducted at CCIPD.
“Previous work involving the radiomic analysis of lung cancer on CT focused on investigating the suspicious lung nodule alone,” she says.
This study differed from previous ones in that researchers evaluated the radiomic features of the nodule in question and an immediate region of the lung parenchyma outside the nodule. Driving the hypothesis were facts already established about tumor-infiltrating lymphocytes and tumor-associated stromal macrophages in the stroma around tumors, which display patterns associated with the likelihood of malignancy that are evident on histopathology.
“We wanted to explore the idea of potentially capturing these patterns on CT imaging, via radiomics,” Beig says.
Beig explains that the research team developed a supervised machine learning pipeline to solve this problem. The data set included 290 patients. A training subset of 145 were benign granulomas. The rest were malignant NSCLC nodules.
“Within the training subset, with equal representation of two classes, we extracted radiomic features from the nodule and the region immediate of the lung parenchyma around the nodule,” she describes.
Following feature selection, a support vector machine classifier was built. This classifier was validated on an independent test set.
“We found that a combination of radiomic patterns of heterogeneity within and outside the tumor could distinguish benign nodules from non–small cell lung cancer on CT scans with 80% accuracy,” Beig reports. “On the same test set, an accuracy of 75% was obtained when we assessed only the nodule by itself.”
Beig says the main lessons were that a great deal of informative signal is present outside the nodule, potentially capturing the malignant microenvironment of a tumor, and combining the radiomic features from the region outside the suspicious nodule with those of the nodule itself can help build more robust radiomic algorithms. She says these findings can have an impact on diagnosis and treatment.
“Adenocarcinomas are the most prevalent subtype of non–small cell lung cancer, making it the most common true-positive finding in a given noncontrast lung cancer screening population,” Beig says. “Granulomas represent the most common and possibly most confounding false-positive finding.”
The study findings generated tremendous interest, given the similar appearance of the two pathologic conditions on imaging. The possibility of a simple and inexpensive diagnostic tool, compared with an existing CT scan, could translate into a better determination of whether a patient will need more invasive and expensive procedures.
“Many patients end up undergoing a surgical intervention in the form of a biopsy, bronchoscopy, or surgical wedge resection for histopathologic confirmation of presence or absence of a malignancy,” Madabhushi says. “Even when no biopsy is recommended, most of these patients undergo multiple CT scans for continued evaluation of the nodule, which leads to unnecessary and potentially harmful radiation.” A radiomic marker that can distinguish between benign and malignant nodules has the potential to enhance diagnostic capabilities as well as reduce patient anxiety and costs.
“Our unique approach was to analyze the CT imaging using AI algorithms to capture subvisual textural clues of the tumor biology that can’t be appreciated by the naked human eye of an expert,” Beig explains. Findings suggested that there is a significant wealth of useful information in the peritumoral region of the tumor. “This region potentially possesses valuable malignancy-related information, such as angiogenic activity that manifests within the peritumoral region on CT imaging,” she says.
Another important aspect of this study, the researchers indicate, was the qualitative assessment of representative patient histology, to unravel the morphometric and biological basis for the most predictive radiomic features. The team found that the immediate vicinity within 5 mm outside the tumor had a unique radiomic signature in adenocarcinomas. In the representative hematoxylin and eosin-stained images, the interface of the tumor had a “rim” of densely packed, tumor-infiltrating lymphocytes and tumor-associated macrophages.
“We hypothesize that this densely packed stromal tumor-infiltrating lymphocytes around adenocarcinomas manifest as smooth texture on CT images and potentially results in this distinctive radiomic signature,” according to the authors.
Remaining Challenges
While clinicians seem to be enthusiastic about the promise of this emerging technology, the promise isn’t completely fulfilled. Those involved in recent research concede that challenges remain. One research challenge involves the issue of where the data are collected. Khorrami says it’s important to use patients from a single institution.
“Accuracy can decrease across multiple sites,” he explains. “When we include patients from multiple sites, the radiomics will change because the acquisition parameters of different CT scanners will change.”
Madabhushi points to the problem of variability as a limitation. “There is a lack of reproducibility of the biomarkers that are identified from the various research-based studies,” he says. “Variability of acquisition parameters, such as contrast enhancement, slice thickness, and convolution kernels, are known to affect the diagnostic performance of radiomic biomarkers.”
Other challenges include interobserver variability and a current lack of standardization. “Radiomic studies often use automatic and semiautomatic methods for segmentation, but since this process is not standardized among different groups of researchers, it should be kept in mind that the reproducibility may also be affected by segmentation,” Madabhushi points out. Also, “findings are mostly based on retrospective analysis where selectively choosing or excluding participants may conflict with the generalizability of the imaging-based radiomic biomarker that is developed. Thus, care should be taken to ensure that the training cohort is not skewed and has balanced classes of different groups in question.”
Finally, while correlations may be derived using carefully controlled experiments, causality is difficult to establish in radiomics. “Ultimately, definitive validation of radiomics needs to be done in a prospective setting in order to truly establish the patient-centric and economic benefits,” Madabhushi says.
However, he also adds that “radiomics still provide a certain degree of interpretability that more black-box approaches such as deep learning are unable to provide. From a clinical adoption perspective, especially when it comes to making recommendations regarding therapy, interpretability of the classifier will be critical.”
Ongoing Research
More research will be required before the promise of radiomics is fulfilled. “Much work remains to be done before its realization as a cancer screening tool that can be integrated into a clinical practice,” Beig says. “[Radiomics] will require careful planning and validation on a larger multisite data set to integrate the human and machine interpretations together in decision support mode.”
Khorrami is at the vanguard of such efforts. He and his colleagues are at work on another study. “In a new project, we want to see if advanced metastatic patients with lung cancer will benefit from immunotherapy or not by having access to baseline CT images,” he reports.
Looking forward, radiomics—while it arose from oncology—should be a technique that can be applied to other diseases or conditions that can be imaged. Still, Madabhushi is quick to note a critical point: Radiomics is not something that will diminish the role of radiologists in diagnosing diseases.
“Rather, radiomics is meant to help radiologists diagnose and characterize disease with greater precision, and it may also help other physicians individualize treatments for their patients,” he says. “Radiomics may play a powerful role in the future management of patients.”
— Dan Harvey is a freelance writer based in Wilmington, Delaware.
