Predicting delayed remission in Cushing's disease using radiomics models: a multi-center study.

Area of Science:

• Endocrinology
• Neurosurgery
• Radiology
• Artificial Intelligence

Background:

• Cushing's disease (CD) management involves transsphenoidal surgery (TSS).
• Predicting delayed remission (DR) post-TSS is crucial for patient outcomes.
• No existing multi-center radiomics models predict DR after TSS in CD.

Purpose of the Study:

• To develop and validate clinical and radiomics models for predicting DR after TSS in CD.
• To leverage multi-center data for robust model development.
• To identify key predictors of delayed remission.

Main Methods:

• Retrospective analysis of 122 CD patients from three centers (2000-2019).
• Utilized T1-weighted gadolinium-enhanced MRI and clinical data.
• Developed automated region of interest (ROI) definition using deep learning.
• Constructed 10 machine learning models, evaluating performance with ROC AUC.

Main Results:

• Overall DR rate was 44.3%.
• Higher BMI and lower postoperative cortisol levels correlated with higher DR rates.
• XGBoost model demonstrated superior performance (AUC 0.767 clinical, 0.819 radiomics).
• SHAP and LIME analyses identified postoperative cortisol and BMI as most significant predictors.

Conclusions:

• Radiomics models offer a promising non-invasive method for predicting DR post-TSS in CD.
• These models can assist neurosurgeons in therapeutic planning.
• Further validation in larger cohorts is warranted.