Qu Hang, Ban Qiqi, Zhou LiangXue, Duan HaiHan, Wang Wei, Peng AiJun
Department of Radiology, Affiliated Hospital of Yangzhou University, Yangzhou University, Yangzhou, Jiangsu province, 225000, China.
Department of Neurosurgery, West China Hospital, Sichuan University, Chengdu, Sichuan province, 610065, China.
BMC Med Imaging. 2025 May 3;25(1):147. doi: 10.1186/s12880-025-01690-5.
Pituitary adenomas (PAs), craniopharyngiomas (CRs), Rathke's cleft cysts (RCCs), and tuberculum sellar meningiomas (TSMs) are common sellar region lesions with similar imaging characteristics, making differential diagnosis challenging. This study aims to develop and evaluate machine learning models using MRI-based radiomics features to differentiate these lesions.
Two hundred and fifty-eight pathologically diagnosed sellar region lesions, including 54 TSMs, 81 CRs, 61 RCCs and 63 PAs, were retrospectively studied. All patients underwent conventional MR examinations. Feature extraction and data normalization and balance were performed. Extreme gradient boosting (XGBoost), support vector machine (SVM), and logistic regression (LR) models were trained with the radiomics features. Five-fold cross-validation was used to evaluate model performance.
The XGBoost model showed better performance than the SVM and LR models built from contrast-enhanced T1-weighted MRI features (balanced accuracy 0.83, 0.77, 0.75; AUC 0.956, 0.938, 0.929, respectively). Additionally, these models demonstrated significant differences in sensitivity (P = 0.032) and specificity (P = 0.045). The performance of the XGBoost model was superior to that of the SVM and LR models in differentiating sellar region lesions by using contrast-enhanced T1-weighted MRI features.
The proposed model has the potential to improve the diagnostic accuracy in differentiating sellar region lesions.
垂体腺瘤（PAs）、颅咽管瘤（CRs）、拉克氏囊肿（RCCs）和鞍结节脑膜瘤（TSMs）是常见的鞍区病变，具有相似的影像学特征，这使得鉴别诊断具有挑战性。本研究旨在开发和评估基于MRI的放射组学特征的机器学习模型，以区分这些病变。
回顾性研究了258例经病理诊断的鞍区病变，包括54例TSMs、81例CRs、61例RCCs和63例PAs。所有患者均接受了常规MR检查。进行了特征提取以及数据归一化和平衡处理。使用放射组学特征对极端梯度提升（XGBoost）、支持向量机（SVM）和逻辑回归（LR）模型进行训练。采用五折交叉验证来评估模型性能。
XGBoost模型表现优于基于对比增强T1加权MRI特征构建的SVM和LR模型（平衡准确率分别为0.83、0.77、0.75；AUC分别为0.956、0.938、0.929）。此外，这些模型在敏感性（P = 0.032）和特异性（P = 0.045）方面存在显著差异。在利用对比增强T1加权MRI特征区分鞍区病变方面，XGBoost模型的性能优于SVM和LR模型。
所提出的模型有潜力提高鞍区病变鉴别诊断的准确性。
