Multiple instance learning for differentiating neoplastic and non-neoplastic brain lesions using MR spectroscopy

Objective To evaluate whether magnetic resonance spectroscopy (MRS) combined with multiple instance learning (MIL) can differentiate neoplastic from non-neoplastic brain lesions by analyzing full spectral patterns rather than isolated metabolite peak ratios. We hypothesized that treating the full spectrum as a metabolic fingerprint, captured through deep learning, would enable more robust tissue classification than conventional single-peak analysis. Methods We retrospectively analyzed MRS data from 422 patients (7442 spectra total): 6598 spectra for training and 844 spectra in the independent internal hold-out test set. Classification was formulated as a MIL problem to address patient-level weak supervision and variable bag sizes. A deep neural network was trained with data augmentation and evaluated with two permutation-invariant aggregation strategies: min–max–mean pooling and attention-based pooling. Diagnostic performance was benchmarked against experienced neuroradiologists. Results The best-performing MIL configuration, combining a convolutional neural network (CNN) with 3Pool aggregation and data augmentation (CNN-3Pool + DA), achieved an AUC of 0.82 with a sensitivity of 0.76 and specificity of 0.80 on the independent hold-out test set. In the MRS-only reader comparison, the participating neuroradiologists showed an average sensitivity of 0.54 and specificity of 0.88. Attention-based MIL additionally enabled visualization of the relative contribution of individual spectra, highlighting clinically established spectral features, including lipid and choline peaks, as well as the Glx region (2.2–2.4 ppm). Conclusions Formulating MRS-based brain tumor classification as a MIL problem enables robust learning from heterogeneous, weakly labeled clinical data. The attention mechanism provides interpretable insights into spectral decision-making and highlights candidate biomarkers for future validation. This framework offers a scalable and interpretable approach to clinical MRS analysis.