Editorial for “Radiomics Nomograms Based on Non‐enhanced MRI and Clinical Risk Factors for the Differentiation of Chondrosarcoma from Enchondroma”

Abstract

Editorial for “Radiomics Nomograms Based on Non-enhanced MRI and Clinical Risk Factors for the Differentiation of Chondrosarcoma from Enchondroma” A radiologic diagnosis in routine practice is always based on a qualitative image interpretation. The images are used only once for the diagnosis and possibly again to compare with the images after treatment to evaluate the response. In the past, these images were neglected and the information was wasted. But today, medical images are considered to be big data sources, which can be quantitatively analyzed under a process called radiomics. If all radiologic data from each institution were shared to a central radiomics database, all radiologic studies could become a valuable and useful source of information. The process of radiomics starts from the method of image acquisition. A universal standard for acquisition of data and reconstruction protocol has been proposed to make the information homogeneous enough for sharing. If the target or interested lesion is identified, segmentation can be done manually, semiautomatically, or autonomically to isolate the region of interest for further data extraction. Each domain of the extracted data is called features, such as size, signal intensity, and textures. A large number of features can be extracted; however, some features are not relevant. Therefore, the appropriate features can be selected for further evaluation. Then, the features can be combined with other available data, such as clinical background, genomics, or biologic markers information, to create variable diagnostic, predictive, and prognostic classification models. At least 10 examples are needed for each feature in a model; therefore, more data provide more power. Radiomics does not only consider the tumor as a single whole volume but also sub-volumes of interest or its different habitats that reflect different tissue properties, in turn a different prognosis. The application of radiomics is a broad spectrum that is well established in oncology. It could be applied from the front to the backend of standard care to support a clinical diagnosis, plan for treatment, prognosis prediction, or even to monitor the disease. Among the subspecialties, not many radiomics studies have been applied in the area of musculoskeletal disorders. However, due to the availability of high computation capabilities and rapidly developing artificial intelligence in the last few decades, radiomics has attracted more musculoskeletal research. To preoperatively differentiate sacral tumors (ie, chordoma, giant cell tumor, and metastasis), Yin et al developed and validated radiomics models based on computed tomography and MRI as well as combined CT and MRI with a certain level of radiomics performance. In this issue of the JMRI, Pan et al used clinical radiomics models that combined clinical tumor location with nonenhanced MRI based on T1-weighted images (T1WI) and fat suppressed T2-weighted images (T2WI-FS) to differentiate enchondroma and chondrosarcoma. This is a common practice for most radiologists in which the decision leads to clinical management. Clinical radiomics models were proven to have high performance in predicting the risk of malignancy in chondroid tumor and performed better than radiologists. This publication reinforces the advantage of a combination of the clinical context with radiomics to enhance the predictive and prognostic performance. This study attempts to solve common clinical questions that impact clinical decisions and management. Radiomics features extracted from contrast-enhanced T1WI may predict overall survival in patients with soft tissue sarcoma and even improve prediction if combined with clinical features according to a study by Spraker et al. Yin et al showed good performance of clinical radiomics as a useful tool for prognostic evaluation to estimate early recurrent pelvic chondrosarcoma. Adopting radiomics for clinical application is time-consuming, especially if the radiomics studies are done retrospectively. The mean time to perform a radiomics analysis to detect local recurrence of soft tissue sarcoma was at least 5 hours per case based on a study by Tagliafico et al. To accelerate the workflow, machine learning might have a promising role in many processes including segmentation. Segmentation is quite crucial and needs to be reproducible since all subsequent information used in an analysis is derived