Differentiating between multiple myeloma (MM) and bone metastases from other cancers can be difficult, but radiomics-based models have the potential to improve diagnostic accuracy.
Differentiating between multiple myeloma of the spine (MM) and bone metastases (BM) related to other types of cancer can be difficult given their sites of occurrence, clinical features, and characteristics of similar imagery. To aid traditional classification methods, a study published in Frontiers in Medicine evaluated new radiomics models based on 18F-fluorodeoxyglucose/CT (18F-FDG PET/CT) positron emission tomography for the classification of MM and BM of the spine.
Misclassification of bone marrow lesions can affect patient survival and quality of life. Misclassified injuries are also often misdiagnosed like other orthopedic diseases, and treatment options are significantly different for each condition. Reducing the risk of misdiagnosis in this context is therefore a crucial effort.
Although serum markers such as creatinine, globulin, and alkaline phosphatase can help differentiate BM from MM, patients with some forms of myeloma tend to have normal or low numbers of these markers. With 18F-FDG PET/CT imaging, the imaging method recommended by the International Myeloma Working Group for MM, both anatomical and metabolic information are used to assess bone damage and lesions with sensitivity and high specificity. But some lesions, such as osteolytic lesions, are still difficult to identify, even for experienced clinicians.
Radiomics uses machine learning to convert imaging features into high-dimensional data, enabling noninvasive assessment of a tumor’s spatial heterogeneity and helping to personalize treatment for each patient. The current study explores the potential of using radiomics in combination with 18F-FDG PET/CT to identify MM versus BM, given that previous research has primarily evaluated radiomics based on CT and MRI images.
A total of 131 patients were included in the study, 86 with a diagnosis of BM and 45 with confirmed MM. A total of 184 lesions were randomly assigned to a training group and a validation group at a ratio of 7:3 to develop the radiomic models. The training group included 80 BM lesions and 49 MM lesions, and the validation group contained 34 and 21 lesions, respectively.
Ten and eight texture features were selected from CT and PET, respectively, to construct the models after absolute least shrinkage and selection operator regression and 10-fold cross-validation. There were 3 radiomics models: 2 constructed with CT and PET and using multivariate logistic regression and a ComModel using PET plus the maximum standardized uptake value of each lesion. Two experienced physicians evaluated the images in a double-blind format to test accuracy against radiomics systems.
In the training and evaluation groups, all 3 radiomics models performed well. The area under the receiver operating characteristic curve (AUC) was 0.909 in the CT training group, 0.949 in the PET training group, and 0.973 in the ComModel training group, and the AUCs of the CT validation group, PET and ComModel were 0.897, 0.929, and 0.948, respectively.
PET and ComModel were significantly better at diagnosing BM and MM compared to expert clinicians, while there was no statistically significant difference between CT model and physician assessment.
The study was limited due to its single center nature. Therefore, further research would help determine the generalizability of the findings. Some patients also had no pathology findings and received their diagnosis based on combined pathology findings and follow-up results, the authors noted. But overall, the study results are promising.
The authors concluded, “Radiomics could transcend subjective visual assessment to provide an objective assessment of lesion and tissue heterogeneity, which served as a novel tool to provide valuable information about the lesion microenvironment that cannot be seen by human eyes.
Jin Z, Wang Y, Wang Y, Mao Y, Zhang F, Yu J. Application of radiomics based on 18F-FDG PET-CT images to identify vertebral multiple myeloma and bone metastases. Front Med (Lausanne). Published online April 18, 2022. doi:10.3389/fmed.2022.874847