SEED-Net: statistical echo evidence and detail modeling for ultrasound lesion classification across multiple organ-specific datasets.

Objective: Discriminative information in ultrasound images is often manifested as local boundary changes, internal echo heterogeneity and differences in fine-scale structure. However, it is susceptible to interference from speckle and uncorrelated background responses. This paper proposed the Statistical Echo Evidence and Detail Network (SEED-Net) for ultrasound lesion classification across multiple organ-specific datasets. Methods: SEED-Net comprises Foreground Tissue Gate, Guided Detail Unit, Channel Recalibration and Statistical Echo Evidence Head, which are used, respectively, for modulating the response of foreground tissue, modelling residuals of local structure and detail, recalibrating channel features, and fusing statistical evidence from medium and deep layers. Each prior branch uses a near-zero gate so that its initial contribution to the residual backbone is approximately zero. The model was trained on three ultrasound datasets, TN5000, SMC-LUD and PU2756, for the classification of ultrasound images of thyroid nodules, liver lesions and lung tumours, respectively. All experiments utilised whole-image input, random initialisation and independent training with five random seeds, and the model was compared with convolutional networks, visual transformers, hybrid architectures and visual state-space models. Results: SEED-Net achieved accuracy values of 0.8702 ± 0.0055, 0.9851 ± 0.0075, and 0.6054 ± 0.0086 on TN5000, SMC-LUD, and PU2756, respectively, with corresponding Macro F1-scores of 0.8384 ± 0.0048, 0.9850 ± 0.0075, and 0.5785 ± 0.0229. It achieved the highest or joint-highest Macro F1-score among the evaluated learning-based models on all three datasets, although its absolute performance on PU2756 remained limited. Ablation experiments demonstrate that the modelling of local details in the backbone and the statistical evidence classification head play complementary roles. Grad-CAM visualisations showed more compact activation patterns around lesion-related regions in the displayed examples, providing qualitative observations of the model responses. Conclusion: By integrating local detail modelling with multi-layer statistical evidence, SEED-Net achieves a favourable balance between performance and computational complexity in ultrasound classification tasks across different organs and varying levels of difficulty.