AI and Cancer Research ── Brain ── 2026-11-14
Brain tumours: treatment, trials and AI
- Abstract
BACKGROUND: O-(2-[18F]fluoroethyl)-L-tyrosine positron emission tomography ([ 18 F]FET PET) is clinically useful in neuro-oncology but limited by cost and availability. We assessed the clinical utility of magnetic resonance imaging (MRI)-derived synthetic [ 18 F]FET PET. METHODS: A total of 489 patients with histologically confirmed gliomas or brain metastases were retrospectively analyzed. A deep learning model generated synthetic PET images from multimodal MRI in the overall cohort and subgroups. Synthetic PET was evaluated against reference PET for image similarity by peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), standardized uptake value ratio (SUVR) analysis by Pearson correlation and Bland-Altman analysis, and blinded reader assessment for glioma grading and high-grade glioma (HGG) versus solitary brain metastasis (SBM) differentiation using receiver operating characteristic analysis with area under the curve (AUC). RESULTS: Synthetic PET showed high similarity to reference PET in the overall, glioma, and metastasis groups (PSNR, 34.7-35.5 dB; SSIM, 0.90-0.96), and SUVR correlations were strong (r = 0.77-0.93, all P 0.05). CONCLUSIONS: MRI-derived synthetic [ 18 F]FET PET may enhance glioma grading and HGG-SBM differentiation, providing a PET-like alternative when FET PET is unavailable.
Journal IF-equivalent: 2.8 (OpenAlex 2-year mean citedness, value as of 2026-11-12, retrieved 2026-11-13; not the official Clarivate IF)Reference: Liang X, Yao C, Diao H, Xiao K, Li F, Lu J. MRI-derived synthetic [¹⁸F]FET PET in neuro-oncology: evaluation in glioma grading and differentiation from solitary brain metastasis. Neuroradiology. 2026 Nov 11 [Epub ahead of print]. doi:10.1007/s00234-026-04266-4. PMID: 43015482.Checked: Abstract only - Abstract
Background and objectivePreoperative identification of fibrous meningioma may help inform surgical planning.We compared contrast-enhanced T1-weighted imaging (T1CE) and fluid-attenuated inversion recovery (FLAIR) radiomics, TabPFN and voxel-input three-dimensional convolutional neural networks (CNNs), and imaging-clinical candidate models.We also evaluated training-only controls and regional-clinical fusion. MethodsThis single-center retrospective study screened 529 records from May 2018 to September 2026 and included 64 cases of World Health Organization (WHO) grade 1 meningioma (18 fibrous and 46 non-fibrous), of which 44 were used for training and 20 for internal evaluation.Automated segmentation and manual quality control defined the tumor and interface regions.In group-aware validation, clinical-only and matched-input classifiers, probability ensembles, and BIG-Huang regional fusion all used a fixed 44/20 split, and all preprocessing and selection were confined to the training data.Performance was assessed at a threshold of 0.5 with 2,000 outcome-stratified bootstrap resamples.Agreement between repeated volume measurements was assessed in 30 cases. ResultsIn the sequence/CNN benchmark, the six-feature and two-feature imaging-clinical candidate models had areas under the curve (AUCs) of 0.810 (95% conditional confidence interval (CI): 0.524-1.000)and 0.905 (0.726-1.000), respectively.Both models had true negatives (TN) = 12, false positives (FP) = 2, false negatives (FN) = 2, and true positives (TP) = 4.In group-aware validation, the five-feature combined TabPFN model had an AUC of 0.726 (0.464-0.929).In the structurally matched BIG-Huang TabPFN regional/fusion models, adding nine clinical variables increased the AUC from 0.702 (0.429-0.905) to 0.833 (0.619-0.976) and reduced the Brier score from 0.234 to 0.151.The 11-input candidate model had an outer out-of-fold (OOF) AUC of 0.422 (0.231-0.625) in the training set.The intra-observer and inter-observer volume intraclass correlation coefficients (ICCs) were 0.9836 and 0.9655, respectively. ConclusionsThis study describes the complementarity of MRI and clinical information through direct-input and regional-score models.The BIG-Huang analysis supports further investigation of regional-clinical integration.Model probabilities require independent calibration assessment before they are used in individual clinical risk communication.
Journal IF-equivalent: 1.2 (OpenAlex 2-year mean citedness, value as of 2026-10-25, retrieved 2026-10-26; not the official Clarivate IF)Reference: Zhu F, Ma H, Tang S, Zou S, Ma H, Guo H. Preoperative Prediction of Fibrous Meningioma Using Contrast-Enhanced T1-Weighted Radiomics and a Tabular Foundation Model: A Single-Center Retrospective Study. Cureus. 2026 Nov 10 [Epub ahead of print]. doi:10.7759/cureus.119408.Checked: Abstract only