A Confidence-Weighted Rule-Based Framework for Multimodal Brain Lesion Classification Using MRI and MRS

Differentiating brain tumors from tumor-like lesions is a persistent clinical challenge due to their overlapping imaging features on conventional radiological scans.Tumor-like lesions such as demyelinating diseases, infections, or post-traumatic changes often mimic neoplastic growths in appearance, leading to potential misdiagnosis and inappropriate treatment decisions.To address this issue, we propose a novel machine-learning-based diagnostic framework that integrates Magnetic Resonance Spectroscopy (MRS) and structural Magnetic Resonance Imaging (MRI) through a confidence-weighted fusion strategy: Final Diagnosis = 0.7 MRS + 0.3 MRI.This weighting reflects the higher metabolic specificity of MRS, while retaining MRI's anatomical detail.Each modality is processed through a specialized pipeline.The MRS pipeline involves image-to-numeric transformation, noise filtering, metabolite concentration-based feature extraction, expertguided feature selection, and a rule-based classifier.The MRI pipeline includes skull stripping, a novel Dynamic Image Thresholding method, multidimensional feature extraction (statistical, volumetric, shape-based), and correlation-based feature selection with a rule-based classifier.Our integrated system achieved 90% classification accuracy on a clinically validated dataset, effectively distinguishing between tumors and tumor-like lesions.Despite the small dataset (n=50) from a single center, stratified cross-validation yielded consistent results (90% accuracy), demonstrating robustness.Future external validation is planned.By introducing a confidence-informed multimodal fusion strategy, the framework provides both high diagnostic accuracy and interpretability, supporting more reliable and informed clinical decision-making in neuro-oncology.