Calibrated ROI-gated conditional computation for high-throughput and backbone-agnostic brain tumor MRI classification

Multi-class brain tumor classification from magnetic resonance imaging must achieve high diagnostic accuracy while maintaining low inference latency and reliable confidence for real-time clinical deployment. High-capacity deep learning models are computationally expensive, and naive acceleration may discard important diagnostic information and reduce confidence reliability. This study proposes a trainable, backbone-agnostic efficiency layer that reallocates computation toward diagnostically relevant regions while preserving global image context and confidence reliability.

The proposed framework learns differentiable spatial localization to identify candidate regions of interest and performs compute-controlled selection to retain informative regions for classification. This enables conditional computation without requiring external segmentation. The method was evaluated on a revised multi-source corpus constructed from three public brain tumor MRI datasets, including the Fernando Feltrin collection, the Nickparvar dataset, and the BRISC dataset, after harmonization of the classification images used in this study. Experiments were conducted across three magnetic resonance imaging modalities (T1, contrast-enhanced T1, and T2) and multiple convolutional neural network backbones using a controlled inference protocol with consistent hardware and batch settings.

The proposed method achieved consistent efficiency improvements across modalities and backbones. Inference throughput increased by 2.3-5.7 times while maintaining classification accuracy comparable to strong baseline models. Calibration analysis showed that removing the calibration component increased expected calibration error from 0.0100 to 0.0535, indicating reduced confidence reliability. Visual analysis confirmed that the model consistently focused on clinically relevant tumor regions.

The proposed framework improves efficiency while preserving diagnostic accuracy and confidence reliability. By combining adaptive region selection, compute control, and calibration, the method enables fast and trustworthy brain tumor classification suitable for resource-constrained and real-time clinical environments.