An MRI-based decision tree to distinguish lipomas and lipoma variants from well-differentiated liposarcoma of the extremity and superficial trunk: Classification and regression tree (CART) analysis

•A new classification tree model was developed using CART analysis.•The first partitioning predictor was based on the presence of contrast enhancement.•The maximal tumor size of 12.75 cm was the second partitioning predictor.•The generalizability of the model was confirmed with external validation. To build and validate a decision tree model using classification and regression tree (CART) analysis to distinguish lipoma and lipoma variants from well-differentiated liposarcoma of the extremities and superficial trunk. This retrospective study included patients who underwent surgical resection and preoperative contrast-enhanced MR imaging for lipoma, lipoma variants, and well-differentiated liposarcoma in two tertiary referral centers. Six MRI findings (tumor size, anatomical location, tumor depth, shape, enhancement pattern, and presence of intermingled muscle fibers) and two demographic factors (patient age and sex) were assessed to build a classification tree using CART analysis with minimal error cross-validation pruning based on a complexity parameter. The model building cohort consisted of 231 patients (186 lipoma and lipoma variants and 45 well-differentiated liposarcoma) from one center, while the validation cohort consisted of 157 patients (136 lipoma and lipoma variants and 21 well-differentiated liposarcoma) from another center. In the CART analysis, the contrast enhancement pattern (no enhancement or thin septal enhancement versus thick septal, nodular, confluent hazy, or solid enhancement) was the first partitioning predictor, followed by a maximal tumor size of 12.75 cm. The tree model allowed distinction of lipoma and lipoma variants from well-differentiated liposarcoma in both the model building cohort (C-statistics, 0.955; sensitivity 80 %, specificity 94.62 %, accuracy 91.77 %) and the external validation cohort (C-statistics, 0.917; sensitivity 66.67 %, specificity 95.59 %, accuracy 91.72 %). The distinction of lipoma and lipoma variants from well-differentiated liposarcoma can be achieved with the simple classification tree model.