AI and Cancer Research ── Brain ── 2026-11-19
Brain tumours: treatment, trials and AI
- Abstract
Preoperative determination of O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status is of major prognostic and predictive relevance in IDH-wildtype glioblastoma, yet currently depends on invasive tissue sampling. This study investigates deep learning-based strategies for MGMT prediction from multiparametric MRIs, with particular emphasis on robustness and generalizability under heterogeneous, multicenter conditions. A total of 949 glioblastoma cases were analyzed, integrating public benchmark data from the Brain Tumor Segmentation Challenge (BraTS) with multi-institutional private cohorts acquired across diverse scanners and imaging protocols. Four complementary experimental settings were designed to compare reproduced literature baselines, private-only training, whole-brain end-to-end models, and segmentation-guided region-of-interest (ROI) pipelines. Tumor delineation was performed using an nnU-Net framework with expert quality control, and multiple three-dimensional convolutional neural network architectures were evaluated. Model performance was highest on the homogeneous BraTS benchmark, where a whole-brain DenseNet121 achieved an area under the receiver operating characteristic curve of 0.80. However, marked performance deterioration was observed when models were evaluated on independent held-out multicenter test set. In contrast, segmentation-guided ROI-based approaches showed a more stable performance profile in this preliminary independent held-out multicenter test. Explicit lesion localization was associated with reduction in false-positive classifications, suggesting a potential benefit for performance stability across heterogeneous datasets, although validation in external cohorts from institutions not represented during training is required. Collectively, these results demonstrate that upstream methodological choices, especially tumor segmentation and ROI definition, remain critical determinants of performance in MRI-based radiogenomic modeling. Future studies should emphasize prospective multicenter validation, domain-shift harmonization, multimodal data integration, and uncertainty-aware deep learning to enhance robustness and interpretability.
Journal IF-equivalent: 2.1 (OpenAlex 2-year mean citedness, value as of 2026-11-16, retrieved 2026-11-18; not the official Clarivate IF) - Abstract
Accurate, interpretable classification of brain tumours from magnetic resonance imaging (MRI) is central to treatment planning, yet high-performing deep models are often poorly calibrated and rarely tested beyond a single dataset. We present QSVQ-Net, a hybrid quantum-classical framework that couples an EfficientNet-B0 backbone and a complex-valued spiking neural network to a morphology-conditioned variational quantum circuit, in which eighteen morphological and radiomic descriptors parameterise the qubit rotations through a learnable mapping. Evaluation covers five-fold cross-validation with confidence intervals, calibration, a matched multi-seed ablation, intrinsic circuit characterisation, auxiliary localisation for mask-free inference and external validation against classical baselines. On the BRISC-2025 test partition, QSVQ-Net attains 98.8% accuracy with an expected calibration error of 0.006; under five-fold cross-validation it reaches 99.48% (95% CI 98.91–100.0), statistically indistinguishable from EfficientNet-B0 (99.57%; paired t-test p = 0.59). The quantum branch performs within confidence-interval overlap of a parameter-matched classical control, so we claim no quantum accuracy advantage. On an independent mask-free dataset, accuracy falls to 72.6% while classical baselines retain 90–93%, a limitation we trace to the morphology branch. QSVQ-Net thus offers a reproducible, transparently characterised quantum-classical design and a candid account of where morphology-conditioned quantum encoding helps and where it does not.
Journal IF-equivalent: 5.9 (OpenAlex 2-year mean citedness, value as of 2026-11-04, retrieved 2026-11-05; not the official Clarivate IF)Reference: Abinash P, Krishnamoorthy N. QSVQ-Net: A morphology-conditioned spiking–variational quantum network for interpretable brain tumour MRI classification. Sci Rep. 2026 Nov 16 [Epub ahead of print]. doi:10.1038/s41598-026-76936-5.Checked: Abstract only