Radiomic Signature of the Substantia Nigra on Neuromelanin-Sensitive MRI Distinguishes Bipolar II Disorder From...

Early differentiation between bipolar disorder type II (BD-II) and unipolar depression (UD) is critical yet challenging owing to overlapping depressive symptoms. This study aimed to develop and validate a radiomic signature based on substantia nigra (SN) neuromelanin-sensitive magnetic resonance imaging (NM-MRI) for distinguishing BD-II, UD, and healthy controls (HCs).

This secondary analysis enrolled 46 drug-naïve BD-II patients, 38 drug-naïve UD patients, and 42 HCs. A total of 2854 radiomic features were extracted from manually segmented SN regions. A multi-stage feature selection pipeline-including inter-observer reproducibility (ICC > 0.75), false discovery rate-corrected ANOVA, correlation analysis (|r| < 0.8), and least absolute shrinkage and selection operator (LASSO) regression-was used to identify a robust signature. Diagnostic performance of linear support vector machine (SVM) and multinomial logistic regression (LR) was evaluated via leave-one-out cross-validation (LOOCV).

The multi-stage selection yielded a parsimonious 7-feature radiomic signature from 2854 initial candidates. For three-class discrimination (BD-II vs. UD vs. HC), SVM outperformed multinomial LR with higher overall accuracy (69.0% vs. 62.7%) and superior discriminative power. SVM's macro-average AUC was 0.849 (95% CI: 0.797-0.884) and its micro-average AUC was 0.840 (95% CI: 0.797-0.884), compared to LR's macro-average AUC 0.812 (95% CI: 0.770-0.861) and micro-average AUC 0.815 (95% CI: 0.770-0.861). Notably, SVM showed marked advantage in the clinically critical BD-II versus UD classification: AUC 0.812 (95% CI: 0.720-0.905) versus LR's 0.756 (95% CI: 0.625-0.860), along with higher specificity (80.6% vs. 67.7%). The SVM model also effectively distinguished patients from HCs (UD vs. HC: AUC 0.869, 95% CI: 0.784-0.954; BD-II vs. HC: AUC 0.860, 95% CI: 0.781-0.939).

This radiomic signature preliminarily demonstrates potential in distinguishing BD-II, UD, and HCs via radiomic analysis of SN NM-MRI. The SVM model, based on the compact radiomic signature, holds promise as an objective tool for addressing BD-II/UD diagnostic challenges, potentially supporting early intervention.