Abstract
<title>Abstract</title> <p>Preoperative ultrasound differentiation among thyroid follicular carcinoma (FTC), primary thyroid lymphoma (PTL), and follicular adenoma (FTA) is a long-standing clinical challenge. Existing AI methods mainly address binary benign–malignant classification, with insufficient work on fine-grained multi-subtype classification-particularly rare entities such as PTL, which requires chemotherapy rather than surgical resection. This paper proposes a multi-class framework integrating an improved Swin-Transformer with ultrasound radiomics for accurate preoperative differentiation of these three types. It adopts a dual-pathway architecture: a Multi-Scale Feature Aggregation Module (MSFAM) and a Channel–Spatial Dual Attention Module (CSDAM) enhance the Swin-Transformer for deep features, while PyRadiomics extraction with intraclass correlation coefficient (ICC) filtering and LASSO regression yields the radiomics subset; the two are integrated via Adaptive Weighted Fusion (AWF) with learnable element-wise gating before Softmax. A retrospective primary cohort of 1,224 pathologically confirmed patients (268 FTC, 131 PTL, 825 FTA) from one institution was split at the patient level into training (= 856), validation (= 184), and internal test (= 184) sets; an independent external cohort of 260 patients (57 FTC, 28 PTL, 175 FTA) from a second institution on different platforms provided out-of-distribution evaluation. The complete model achieved internal-test ACC 90.76% (95% CI: 86.85–94.20%) and Macro-AUC 0.9487 (95% CI: 0.9214–0.9721), and robust external performance (ACC 87.31% [95% CI: 82.64–91.28%]; Macro-AUC 0.9247 [95% CI: 0.8912–0.9542]). Ablation validated each module's contribution, and comparison against six mainstream methods showed superiority (all < 0.00833 by DeLong test with Bonferroni correction for six pairwise comparisons). In a blinded reader study, the proposed system improved two senior physicians' diagnostic accuracy by an average of 10.60 percentage points (both < 0.001 by paired McNemar's test) and their inter-reader agreement from Cohen's \(\kappa\) = 0.65 to \(\kappa\) = 0.82. After temperature scaling, the model was well-calibrated (Brier 0.081, ECE 0.028), and decision curve analysis confirmed net clinical benefit across the relevant probability threshold range. Grad-CAM + + attention matched pathological features, with a mean IoU of 0.72 against expert-annotated regions of clinical interest on 60 cases, supporting interpretability for trustworthy clinical use.</p>