Spectral–spatial Mamba with uncertainty-guided refinement for thyroid nodule diagnosis

Thyroid nodule malignancy assessment is a high-stakes predictive task in clinical radiology, where diagnostic uncertainty directly affects biopsy referral, surgical planning, and patient outcomes. We present SSMURNet (Spectral–Spatial Mamba Uncertainty-Refined Network), a deep learning framework for confidence-calibrated thyroid nodule classification from ultrasound images. The proposed architecture integrates three complementary components: (i) a dual-branch spectral–spatial encoder combining Fast Fourier Transform (FFT) frequency analysis with Mamba state space modelling to capture microcalcification signatures and long-range morphological dependencies; (ii) an Uncertainty-Guided Attention (UGA) mechanism that enhances feature representation in diagnostically ambiguous regions using epistemic uncertainty estimates; and (iii) a hierarchical evidential learning objective that produces calibrated malignancy probabilities in a single forward pass without Monte Carlo sampling. Evaluated on 7288 pathologically confirmed thyroid ultrasound images, SSMURNet achieved 92.50% accuracy (95% CI: 90.95–93.97%) and 0.9700 AUROC, with sensitivity and specificity of 95.17% and 89.25%, respectively. Calibration analysis yielded a Brier score of 0.1218. Incorrectly classified samples showed higher mean uncertainty (0.6446) than correctly classified samples (0.6192), indicating meaningful uncertainty awareness. External validation on an independent thyroid ultrasound dataset achieved 85.42% accuracy and 0.9185 AUROC without retraining, demonstrating robust cross-domain generalisation. These findings position SSMURNet as a promising algorithmic prototype for thyroid cancer screening, warranting prospective multicenter validation before clinical deployment.