| Summary: | We developed a radiomics model to differentiate hematologic marrow diseases and compared the performance with radiologists' readings and a quantitative measurement. Patients were retrospectively analyzed from the diseased (n = 254) and control groups (n = 230). A sagittal T1-weighted lumbar spinal MR image was normalized by an intervertebral disk, and bone marrow was segmented. A hundred features were extracted, and final features were selected using Principle Component Analysis (PCA) and least absolute shrinkage and selection operator (LASSO). Finally, Random forest (RF) and logistic regression (LR) models were trained. Two radiologists with different levels of experience analyzed the images for the presence of bone marrow diseases, independently. The area under the receiver operating characteristic curves (AUC) and decision curve analysis (DCA) was evaluated. Among the subjects, 363 cases were assigned as a training set and 121 as a validation set. The combination of LASSO and RF produced the best results. With the validation set, the sensitivity (SE) was 87.3%, specificity (SP) was 86.2% and AUC was 0.928 ($p <; 0.05$ ). We selected Firstorder -Maximum as the best feature to identify diseased marrows, which achieved SE of 75.0% and AUC of 0.787 ($p <; 0.05$ ). The reader with 11 years of experience yielded SE of 86.5% and AUC of 0.861 ($p <; 0.05$ ). The second reader with 1 year of experience yielded SE of 75.0% and AUC of 0.767 ($p <; 0.05$ ). We demonstrated the advantage of bone marrow radiomics over conventional methods of diagnosing with radiologists' readings and quantitative measurements.
|
|---|
