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

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Neuroscience

Background:

• Accurate multi-class brain tumor classification from MRI is crucial for clinical deployment.
• High-capacity deep learning models are computationally intensive, posing challenges for real-time applications.
• Existing acceleration methods may compromise diagnostic information and confidence reliability.

Purpose of the Study:

• To develop a trainable, backbone-agnostic efficiency layer for brain tumor classification.
• To reallocate computation towards diagnostically relevant regions while preserving global context and confidence.
• To enable efficient and reliable real-time clinical deployment of deep learning models.

Main Methods:

• A differentiable spatial localization approach identifies regions of interest.
• Compute-controlled selection retains informative regions, enabling conditional computation.
• The framework was evaluated on harmonized multi-source MRI datasets across different modalities and CNN backbones.

Main Results:

• Consistent efficiency gains (2.3-5.7x throughput increase) were observed across modalities and backbones.
• Classification accuracy was maintained comparable to strong baseline models.
• Calibration analysis confirmed preserved confidence reliability, with minimal increase in expected calibration error.

Conclusions:

• The proposed framework enhances efficiency for brain tumor classification without sacrificing diagnostic accuracy or confidence.
• Adaptive region selection, compute control, and calibration ensure fast and trustworthy classification.
• This method is suitable for resource-constrained and real-time clinical environments.