A Verifiable Framework for Brain Tumor Classification: Combining Vision Transformers, Class-Weighted Learning, and...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neuro-oncology
• Machine Learning in Healthcare

Background:

• Automated brain tumor classification from single MRI slices is challenging due to limited context.
• Existing methods struggle with the nuances of single post-contrast axial T1-weighted slices.

Purpose of the Study:

• To develop and evaluate a novel slice-level classification framework for brain tumors.
• To assess the performance of a Swin-Tiny Transformer model in this task.
• To incorporate post hoc logical consistency checks for enhanced reliability.

Main Methods:

• A four-class classification framework using a fine-tuned Swin-Tiny Transformer.
• Incorporation of inverse-frequency class-weighted learning.
• A prototype symbolic model theory (SMT)-based symbolic auditing layer for logical consistency.

Main Results:

• Achieved 97.42% slice-level accuracy on an internal dataset, outperforming convolutional baselines.
• Demonstrated robust performance on an independent dataset (94.82% accuracy) despite distribution shifts.
• The symbolic auditing layer identified a small percentage of constraint-violating predictions.

Conclusions:

• Hierarchical shifted-window attention is valuable for slice-level MRI classification.
• The proposed framework shows promise for neuro-oncologic imaging, with potential for clinical deployment after further validation.
• The study provides an empirical benchmark and a prototype for logical auditing in medical imaging AI.