Machine learning models for differential diagnosis of lipomas and soft-tissue liposarcomas: a multicenter cross-sectional study
Digital Diagnostics · 2026-09-03 · 7 (3) · 261-284
DOI: 10.17816/dd695849 ↗Abstract
BACKGROUND: Differences in management strategies, therapeutic approaches, and overall prognosis between soft-tissue lipomas and liposarcomas underscore the clinical necessity for accurate differentiation. However, differential diagnosis remains inherently challenging due to overlapping semiotic features on magnetic resonance imaging. Consequently, objective and automated analytical methods for magnetic resonance imaging interpretation are required. AIM: To develop machine learning models incorporating semiotic and radiomic features and compare their diagnostic efficacy in distinguishing soft-tissue lipomas from liposarcomas. METHODS: A retrospective multicenter cross-sectional study was conducted across three research centers from 2012 to 2025. Inclusion criteria were histologically verified lipomas or liposarcomas and a contrast-enhanced or non-contrast magnetic resonance imaging examination performed within one month prior to antitumor therapy. Non-inclusion criteria comprised lack of histological verification, histologically verified diagnosis other than lipoma or liposarcoma, incomplete magnetic resonance imaging acquisition protocols, and pronounced image artifacts precluding interpretation. The cohorts were categorized by diagnosis (lipoma vs. liposarcoma) verified by postoperative histopathology. Predictors included semiotic features and radiomic descriptors extracted from volumetric tumor segmentations on native T1- and T2-weighted sequences. The models were build using seven machine learning algorithms: logistic regression, linear discriminant analysis, support vector machines, random forest, naive Bayes classifier, k-nearest neighbors, and decision tree. RESULTS: Models constructed exclusively with semiotic features demonstrated superior predictive performance and robustness. Following hyperparameter optimization, linear discriminant analysis and naive Bayes models achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.944 [95% CI, 0.810–1.000] and (95% CI, 0.773–1.000) with an overall accuracy of 0.842. Conversely, models relying solely on radiomic features exhibited lower diagnostic performance. Peak pre- and post-optimization ROC-AUC values were 0.833 and 0.611 for the decision tree model based on the features extracted from T1-weighted images, 0.833 and 0.700 for the random forest model (T2-weighted images), and 0.851 and 0.724 for the random forest/logistic regression models (T1 + T2-weighted images). A hybrid model integrating semiotic and radiomic features extracted from T2-weighted images yielded a ROC-AUC of 0.912 (95% CI, 0.744–1.000). Using the full feature array (214 radiomic and 7 semiotic features) in a subset of 88 patients failed to yield sustained accuracy improvements (pre- and post-optimization ROC-AUC of 0.848 and 0.749, respectively). CONCLUSION: The models provide a viable foundation for clinical decision-support systems aimed at refining preoperative evaluation of lipomas and liposarcomas. External validation across larger multicenter cohorts is required to confirm diagnostic generalizability and establish feasibility of implementation into clinical practice.
EUVIMED