Fig 1.
Flowchart of radiomic-based machine learning model development.
In step 1, MR images are manually segmented by attending clinicians. In step 2, feature extraction is performed using a LASSO model. In step 3, classifiers are trained and tested using cross-validation. In step 4, a probability score ranging from 0–1 is produced for each tumor (0 = benign; 1 = malignant).
Table 1.
Demographic and clinical features of the benign and malignant groups.
Fig 2.
Feature importance generated by a logistic regression classifier to differentiate benign and malignant myxoid tumors.
The classifier was fitted using the top 17 radiomic and clinical features, which were selected most frequently by the LASSO models. Features are arranged from left to right in decreasing order of selection frequency. The top 6 features (on the left) were selected by all LASSO models. Normalized value is the log odds ratio of each feature’s association with malignancy, i.e., the greater the magnitude of a positive value, the more likely malignant.
Table 2.
Classification performance of five ML models using T1+T2/STIR images.
Table 3.
Classification performance of radiomics vs. radiologists in distinguishing benign and malignant myxoid tumors using T1+T2/STIR images.
Table 4.
Cumulative accuracy results based on confidence level of radiologist 1 and 2 with three iterations of diagnostic classification.