MRI-based 2.5D deep learning radiomics nomogram for the differentiation of benign versus malignant vertebral compression fractures.

Objective: Vertebral compression fractures (VCFs) represent a prevalent clinical problem, yet distinguishing acute benign variants from malignant pathological fractures constitutes a persistent diagnostic dilemma. To develop and validate a MRI-based nomogram combining clinical and deep learning radiomics (DLR) signatures for the differentiation of benign versus malignant vertebral compression fractures (VCFs). Methods: A retrospective cohort study was conducted involving 234 VCF patients, randomly allocated to training and testing sets at a 7:3 ratio. Radiomics (Rad) features were extracted using traditional Rad techniques, while 2.5-dimensional (2.5D) deep learning (DL) features were obtained using the ResNet50 model. These features were combined through feature fusion to construct deep learning radiomics (DLR) models. Through a feature fusion strategy, this study integrated eight machine learning architectures to construct a predictive framework, ultimately establishing a visualized risk assessment scale based on multimodal data (including clinical indicators and Rad features).The performance of the various models was evaluated using the receiver operating characteristic (ROC) curve. Results: The standalone Rad model using ExtraTrees achieved AUC=0.801 (95%CI:0.693-0.909) in testing, while the DL model an AUC value of 0.805 (95% CI: 0.690-0.921) in the testing cohort. Compared with the Rad model and DL model, the performance superiority of the DLR model was demonstrated. Among all these models, the DLR model that employed ExtraTrees algorithm performed the best, with area under the curve (AUC) values of 0.971 (95% CI: 0.948-0.995) in the training dataset and 0.828 (95% CI: 0.727-0.929) in the testing dataset. The performance of this model was further improved when combined with clinical and MRI features to form the DLR nomogram (DLRN), achieving AUC values of 0.981 (95% CI: 0.964-0.998) in the training dataset and 0.871 (95% CI: 0.786-0.957) in the testing dataset. Conclusion: Our study integrates handcrafted radiomics, 2.5D deep learning features, and clinical data into a nomogram (DLRN). This approach not only enhances diagnostic accuracy but also provides superior clinical utility. The novel 2.5D DL framework and comprehensive feature fusion strategy represent significant advancements in the field, offering a robust tool for radiologists to differentiate benign from malignant VCFs.