Radiomics-Based Multiple Machine Learning Approaches for Investigating Medial Wall Invasion of the Cavernous Sinus in Pituitary Adenomas

Yuyang Chen¹Jiansheng Zhong²Pengwei Hou³Xiaoyu Wang⁴Jun Li²Tianshun Feng²Ziqi Li²LIANG FENG WEI²Yuhui Chen^2*Shousen Wang^2*

• ¹Fudan University Huashan Hospital Department of Neurology, Shanghai, China
• ²900th Hospital of the People's Liberation Army Joint Logistic Support Force, Fuzhou, China
• ³Jinjiang Municipal Hospital, Quanzhou, China
• ⁴Changle District People's Hospital, Fuzhou, China

Objective: This study aims to develop a predictive model for cavernous sinus dural invasion in pituitary adenomas by retrospectively analyzing clinical and imaging data. It explores the associations between clinical and radiomics features and cavernous sinus dural invasion. Methods: Clinical data and coronal T2-weighted MRI images were collected from patients diagnosed with pituitary adenomas at our institution between December 2012 and December 2022. Tumor regions of interest (ROIs) were segmented using 3D Slicer, and radiomics features were extracted. Statistically significant radiomics features were identified using Lasso regression and univariate analysis. Clinical features were screened using univariate and multivariate logistic regression analyses. These selected features were incorporated into ten machine learning algorithms to construct three predictive models: a clinical feature model, a radiomics feature model, and a combined clinical and radiomics feature model. Model performance was evaluated to determine the best-performing model, which was further interpreted. Results: A total of 252 patients with histopathologically confirmed pituitary adenomas were included. The analysis identified Knosp grade, tumor left-right diameter, pedunculated satellite tumor, and clival invasion as significant clinical predictors, along with radiomics features including original.4, original.10, log-sigma-5-0-mm-3D.29, log-sigma-5-0-mm-3D.91, wavelet-LLH.37, wavelet-LHL.37, and wavelet-HLL.8. The combined clinical and radiomics model outperformed models based solely on clinical or radiomics features. Among the ten machine learning algorithms, the LightGBM model demonstrated the best predictive performance, achieving an area under the curve (AUC) of 0.86 and an accuracy (ACC) of 0.76. Conclusions: A machine learning model integrating clinical and radiomics features can effectively predict cavernous sinus dural invasion in pituitary adenomas preoperatively, providing a reliable basis for diagnosing tumor invasiveness and developing surgical plans. The LightGBM algorithm exhibited the highest predictive efficacy. Furthermore, the pedunculated satellite tumor feature emerged as a novel imaging marker for cavernous sinus dural invasion, offering new insights into the study of invasive pituitary adenomas.