Noncontrast MRI-based machine learning and radiomics signature can predict the severity of primary lower limb lymphedema.

OBJECTIVE

According to International Lymphology Society guidelines, the severity of lymphedema is determined by the difference in volume between the affected limb and the healthy side divided by the volume of the healthy side. However, this method of measuring volume is time consuming, laborious, and has certain errors in clinical applications. Therefore, this study aims to explore whether machine learning radiomics features based on noncontrast magnetic resonance imaging (MRI) can predict the severity of primary lower limb lymphedema.

METHODS

A retrospective analysis of 119 patients with primary lower limb lymphedema. The enrolled patients were divided into a nonsevere group (mild and moderate) and a severe group. Using the semiautomatic threshold method in ITK-snap software on the patient's noncontrast MRI, we filled the area between the subcutaneous tissue and muscle of the edematous site. The PyRadiomics software package was used to extract radiomic features. The radiomic features were analyzed using the t test or Mann-Whitney test. Subsequently, Pearson correlation testing and least absolute shrinkage and selection operator screening were performed. Using Scikit-learn, the remaining features were used to construct five models: logistic regression, support vector machine, random Forest, ExtraTrees, and light gradient boosting machine. The predictive performance were evaluated by the receiver operating characteristic curve, and the sensitivity and specificity of these measures were calculated. The predictive curve was used to evaluate the performance of the predictive model in guiding decisions for nonsevere and severe lymphedema patients.

RESULTS

The enrolled patients including 28 patients with mild lymphedema (grade I), 38 patients with moderate lymphedema (grade II), and 53 patients with severe lymphedema (grade III) was conducted. A total of 1196 features were extracted, and after Pearson correlation testing and least absolute shrinkage and selection operator screening, 21 nonzero features were selected. The ExtraTree model performed the best, with an area under the curve of 0.974 (95% confidence interval, 0.9437-1.0000) in the training set, a sensitivity of 89.2%, and a specificity of 95.7%. In the test set, these values were 0.938 (95% confidence interval, 0.8539-1.0000), 75%, and 100%, respectively. The decision curve showed that when the predicted probability was between 16% and 78%, the net benefit of the ExtraTree model was greater than that of the two extreme curves, indicating strong clinical value in guiding decisions for nonsevere and severe lymphedema patients.

CONCLUSIONS

All five models performed well in distinguishing between the nonsevere group and the severe group. Noncontrast MRI-based machine learning radiomics signature can predict the severity of primary lower limb lymphedema.