Sci Rep. 2025 Dec 17. 15(1): 44104
Myxoid liposarcoma (MLPS) is a rare soft tissue sarcoma characterized by histopathological variability, which poses challenges for accurate grading and treatment planning. This study evaluated the feasibility of an automated MRI-based pipeline that combines deep learning and radiomics for non-invasive tumor assessment. In a retrospective multicenter cohort of 48 patients with histologically confirmed MLPS, a 3D U-Net convolutional neural network was trained to perform automatic tumor segmentation on axial T2-weighted MR images. Radiomics features were subsequently extracted from the segmented volumes and used to train a Random Forest classifier for predicting tumor grade, defined by centralized histopathological review according to WHO criteria. The segmentation model achieved a median Dice similarity coefficient of 0.892. The radiomics-based grading classifier reached a mean area under the curve of 0.745, with an F1-score of 0.729 and a balanced accuracy of 0.723 in distinguishing high-grade from low-grade tumors. Most classification errors occurred in borderline or histologically heterogeneous cases. These findings suggest that automated segmentation and radiomics analysis may offer valuable support for MLPS grading and complement histopathology, particularly in diagnostically complex cases. Further prospective validation in larger cohorts is warranted.
Keywords: Deep learning; Liposarcoma, myxoid; Magnetic resonance imaging; Neoplasm grading; Sarcoma