Solitary fibrous tumor (SFT) represents a form of invasive soft tissue sarcoma (1). Previously grouped with hemangiopericytomas (HPC) under the term "solitary fibrous tumor/hemangiopericytoma" by the CNS WHO in 2016, the classification was revised in 2021, exclusively designating these lesions as SFT (2, 3). Meningioma, the most common adult intracranial tumor, is categorized into three WHO grades (3, 4). Angiomatous meningioma (AM), a grade 1 meningioma, poses diagnostic challenges owing to its histological resemblance to SFT (5, 6). Unlike AM, which generally has a favorable prognosis following surgical resection, SFT often manifests with extracranial metastasis and local recurrence (6–8). Accurate preoperative differentiation between these tumors is paramount for treatment planning.

Magnetic Resonance Imaging (MRI) is a primary tool for evaluating central nervous system malignancies (9). Contrast-Enhanced T1-Weighted Imaging (CE-T1WI) is instrumental in evaluating blood-brain barrier integrity and delineating tumor characteristics. T2-Weighted Imaging (T2WI) is valuable for superior soft tissue resolution in tumor detection (10). Preoperative imaging methods, including the intratumoral flow void sign and apparent diffusion coefficient (ADC) map, have been explored for differentiating SFT from AM (11, 12). However, these techniques are subjective and heavily reliant on radiologist expertise, underscoring the necessity for more objective and quantitative diagnostic approaches.

Radiomics, an emerging field in medical imaging, leverages data-characterization algorithms to extract quantitative features from radiological images, thereby augmenting prognostic monitoring and treatment strategies in oncology (13). Central to this domain, machine learning, particularly deep learning, plays a critical role in feature analysis, significantly contributing to advancements in medical imaging (14). By objectively assessing tumor heterogeneity, radiomics propels the development of precision oncology (15). Numerous studies have demonstrated that its application spans various cancer types, including stomach and esophageal cancers, and extends to intracranial tumors like gliomas and meningiomas (16, 17). In addition, radiomics has proven effective in differentiating between SFT and AM (5, 18, 19). According to Li et al. (19), the area under the receiver operating characteristic curve (AUC) of the CE-T1WI-based radiomics algorithm for distinguishing SFT and AM was 0.90, significantly higher than the AUCs of three neuroradiologists (AUC=0.69, 0.70, and 0.73). Nevertheless, MRI-based radiomics nomograms, integrating both conventional imaging features and radiomics for differentiating SFT and AM, remain underexplored.

Here, our research aims to ascertain the diagnostic performance of MRI-based radiomics nomogram in preoperative differentiation between SFT and AM, using data from four centers to enhance the precision of therapeutic decision-making.

The preoperative distinction between SFT and AM is of paramount clinical significance. SFT, characterized by its invasive nature, has a propensity for recurrence and metastasis to extracranial organs post-surgical resection. Typically, the primary therapeutic strategy encompasses postoperative radiotherapy or chemotherapy, supplemented by regular clinical follow-ups to monitor patient prognosis (24). Conversely, AM exhibits a benign pathology, with low aggressive growth potential and recurrence rates, often resulting in favorable outcomes after gross total resection (25). SFT and AM exhibit similar MRI findings. Although previous studies have demonstrated that certain MRI features—such as tumor size, signal intensity, vascular flow voids, and the dural tail sign—can aid in differentiating between SFT and AM (26). However, traditional MRI features are susceptible to the influence of physician experience due to the lack of objectivity and quantitative analysis, so differentiating between these two entities using conventional MRI poses a considerable challenge. In our study, we selected the radiomics algorithm with superior predictive performance, integrating it with the clinical model to construct a radiomics nomogram specifically for distinguishing SFT from AM. The radiomics nomogram AUC was 0.989 in the validation set, outperforming the clinical model and the radiomics algorithm in predictive capability. Demonstrating satisfactory calibration and net benefits, the radiomics nomogram appears reliable for differentiating AM from SFT.

Image segmentation in machine learning for intracranial tumors, a crucial step, demands high reproducibility. Current methodologies range from manual to semiautomatic and fully automatic approaches (27, 28). Hu et al. (29) implemented a semiautomatic method based on a signal intensity threshold and edge-based algorithms via 3D Slicer, achieving efficient tumor segmentation. This approach automatically aligned the ROIs with multimodal MRI images, proving more efficient than manual sketching. However, despite its efficiency, automatic mapping encounters challenges, particularly with the variability in tumor morphology. In our study, manual segmentation was meticulously conducted using 3D Slicer to ensure precise tumor delineation.

The efficacy of morphological features in differentiating between AM and SFT remains contentious. Various studies advocate for their diagnostic relevance, noting characteristics typical of SFT such as large volume, irregular shape, uneven enhancement, vascular flow voids, dural stenosis, necrosis, and bone deterioration (19, 30–33). Contrarily, our research suggests a divergence from these findings, potentially due to the subjective nature of clinical observations. Our multivariate logistic regression analyses identified age, width, and dural tail sign as independent predictors in distinguishing SFT from AM. In addition, the AUC of the clinical model in our validation set was inferior to that of the radiomics algorithms and nomograms, underscoring the dependency of the clinical model's performance on radiologist expertise and emphasizing the superior discriminative performance of machine learning algorithms (33).

Radiomics, a constantly evolving new subject in medical imaging, refers to the quantitative extraction of radiological features from two-dimensional or three-dimensional medical images. Its integration with clinical data has proven beneficial in monitoring tumor progression and aiding personalized treatment strategies (34). Machine learning plays a crucial role in analyzing brain tumor MRI imaging data, facilitating deeper exploration of patterns that can aid in diagnosis. At present, domestic and foreign scholars have applied machine learning based on radiomics methods to study SFT and AM. Fan et al. (35) utilized SVM to develop a model for distinguishing SFT from AM using CE-T1WI, T2WI, and a combined CE-T1WI and T2WI sequence. The combined CE-T1WI and T2WI model demonstrated the highest predictive performance, achieving an AUC of 0.90. Kong et al. (36) assessed various algorithms' effectiveness in differentiating SFT from AM using multiparameter MRI sequences. Their results indicated that the performance improvements offered by the different algorithms were limited. However, most studies typically utilize only one machine learning algorithm or a single traditional MRI sequence to construct their models, often using data from a single center. Validating the generalizability of these models necessitates the inclusion of multicenter data. Therefore, in line with previous research, we integrated data from multiple centers and employed LR, SVM, and KNN algorithms in conjunction with conventional MRI sequences to establish radiomics models, and the models ultimately showed robust predictive capabilities. Furthermore, our integrated model demonstrated superior diagnostic accuracy, outperforming the individual LR, SVM, and KNN models, as well as the clinical model. This underscores that while tumor radiomics models possess enhanced predictive abilities compared to the clinical model alone, the integration of clinical data is indispensable. The synergy of these elements effectively differentiates between SFT and AM. A radiomics nomogram, established by Wei et al. to differentiate SFT from AM, showcased remarkable diagnostic capability, with AUCs of 0.985 and 0.917 in the training and validation sets, respectively (37).

In our study, 1,166 radiomics features, predominantly texture features representing second-order attributes that reflect voxel/pixel relationships, were extracted. Texture features illustrate not only pixel intensity distribution, but also how quantized pixels position each other (38). The feature selection process in machine learning, crucial for the reliability of radiomics algorithms, involved the use of mRMR and LASSO. LASSO is particularly effective in datasets with numerous features but limited samples without noticeably increasing the bias, while mRMR facilitates the selection of features with minimal redundancy and maximum relevance (39, 40). Subsequently, seventeen features were selected to develop radiomics models using SVM, LR, and KNN classifiers. SVM is adept at analyzing high-dimensional, small-scale data. KNN excels in processing nonlinear data, identifying multiple predictive biomarkers in clinical datasets. LR, as a classification method, investigates the relationship between specific outcome probabilities and features (41, 42). Our results indicated that LR yielded the highest accuracy and AUC, prompting their adoption for the optimal radiomics model. Integrating this model with the clinical model to create the radiomics nomogram resulted in excellent calibration and accurate differentiation between SFT and AM in the validation set, affirming its efficacy and reliability in diagnosing these conditions.

Multicenter studies, essential for obtaining extensive sample sizes and enhancing classifier generalization, encounter challenges like data heterogeneity and variability arising from different scanners and methodologies. To mitigate these issues, we implemented preprocessing steps such as resampling, denoising, and wavelet transformation. In the validation set, clinical model, radiomics algorithm, and the nomogram demonstrated AUCs of 0.911, 0.968, and 0.989, respectively, suggesting robust generalizability (43).

This investigation had certain restrictions. Firstly, its retrospective nature and relatively small size might introduce selection bias. Considering the retrospective design, scanner and protocol heterogeneity were addressed through resampling to stabilize model performance. Additionally, patient distributions varied between training and validation sets. Lastly, our analysis was confined to images from the CE-T1WI and T2WI. Future studies aim to include a broader range of imaging sequences to further enhance the predictive accuracy and generalizability of the model.

In summary, compared to conventional MRI, the MRI-based radiomics nomogram demonstrates greater efficacy in differentiating SFT and AM, offering significant information for the subsequent treatment and detection. Furthermore, the study of a larger prospective dataset is needed to certify the real value of nomogram.