Descriptive Evidence Synthesis and Methodological Risk-of-Bias Audit of MRI-Based AI Models for Intracranial Tumor Differentiation

This project conducted a comprehensive systematic review and descriptive evidence synthesis evaluating the diagnostic, predictive, and segmentation performance of MRI-based artificial intelligence (AI), radiomics, and deep learning architectures (CNNs, transfer learning, and computer-aided diagnostic models) for preoperative differentiation of sellar, parasellar, skull-base, and extra-axial intracranial lesions across $N = 21$ studies. The target pathology spectrum encompassed pituitary adenomas/PitNETs, craniopharyngiomas, Rathke cleft cysts (RCC), tuberculum sellae meningiomas, solitary fibrous tumors/hemangiopericytomas (SFT/HPC), and vestibular schwannomas. While initial protocol objectives aimed for quantitative pooling, a formal Network Meta-Analysis (NMA) was precluded due to structural evidence disconnects across all 50 extracted model comparisons, alongside widespread under-reporting of $95\%$ confidence intervals, standard errors, and usable $2 \times 2$ contingency matrices in primary literature. Executing a structured descriptive synthesis and QUADAS-2 risk-of-bias audit revealed that 13 of 21 studies ($62\%$) exhibited high risk of bias, driven by single-center retrospective designs, internal-only cross-validation, and potential image-level data leakage in public benchmark datasets. Detailed narrative comparisons were constructed for genuine overlapping clinical tasks, including adamantinomatous vs. papillary craniopharyngioma subtyping (AUCs spanning $0.763$ to $0.890$) and SFT/HPC vs. meningioma differentiation (AUCs up to $0.920$). The synthesis highlighted crucial reporting deficits in neuro-oncological AI research, establishing methodological standards for future multi-center validation and clinical translation.