Prediction of High-Risk Cytogenetic Status in Multiple Myeloma Based on Magnetic Resonance Imaging: Utility of Radiomics and Comparison of Machine Learning Methods

J Magn Reson Imaging. 2021 May 12. doi: 10.1002/jmri.27637. Online ahead of print.

ABSTRACT

BACKGROUND: Radiomics has shown promising results in the diagnosis, efficacy, and prognostic assessments of multiple myeloma (MM). However, little evidence exists on the utility of radiomics in predicting a high-risk cytogenetic (HRC) status in MM.

PURPOSE: To develop and test a magnetic resonance imaging (MRI)-based radiomics model for predicting an HRC status in MM patients.

STUDY TYPE: Retrospective.

POPULATION: Eighty-nine MM patients (HRC [n: 37] and non-HRC [n: 52]).

FIELD STRENGTH/SEQUENCE: A 3.0 T; fast spin-echo (FSE): T1-weighted image (T1WI) and fat-suppression T2WI (FS-T2WI).

ASSESSMENT: Overall, 1409 radiomics features were extracted from each volume of interest drawn by radiologists. Three sequential feature selection steps-variance threshold, SelectKBest, and least absolute shrinkage selection operator-were repeated 10 times with 5-fold cross-validation. Radiomics models were constructed with the top three frequency features of T₁ WI/T₂ WI/two-sequence MRI (T₁ WI and FS-T₂ WI). Radiomics models, clinical data (age and visually assessed MRI pattern), or radiomics combined with clinical data were used with six classifiers to distinguish between HRC and non-HRC statuses. Six classifiers used were support vector machine, random forest, logistic regression (LR), decision tree, k-nearest neighbor, and XGBoost. Model performance was evaluated with area under the curve (AUC) values.

STATISTICAL TESTS: Mann-Whitney U-test, Chi-squared test, Z test, and DeLong method.

RESULTS: The LR classifier performed better than the other classifiers based on different data (AUC: 0.65-0.82; P < 0.05). The two-sequence MRI models performed better than the other data models using different classifiers (AUC: 0.68-0.82; P < 0.05). Thus, the LR two-sequence model yielded the best performance (AUC: 0.82 ± 0.02; sensitivity: 84.1%; specificity: 68.1%; accuracy: 74.7%; P < 0.05).

CONCLUSION: The LR-based machine learning method appears superior to other classifier methods for assessing HRC in MM. Radiomics features based on two-sequence MRI showed good performance in differentiating HRC and non-HRC statuses in MM.

EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 2.

PMID:33979466 | DOI:10.1002/jmri.27637