Multi-Scale ConvNeXt and Late Fusion of Segmentation-Derived Morphological and Echotextural Features for the Benign–Malignant Differentiation of Thyroid Nodules on Ultrasound

Distinguishing between benign and malignant thyroid nodules from ultrasound images is challenging due to low contrast, speckle noise, and indistinct borders. In this study, a gate-free residual late fusion model is presented that combines three-scale ultrasound images with segmentation-derived quantitative morphological and echotextural features. For diagnostic evaluation, the official 3,500/500/1,000 training/validation/test split of the open-access TN5000 dataset, comprising 5,000 images, was preserved. Views at 1.0×, 1.5×, and 2.5× magnification from the annotation box center were generated; a 768-dimensional image representation was extracted using a frozen shared ConvNeXt-Tiny model. Eighty-three morphology-boundary and 85 echodensity-frequency features were calculated from pseudo-masks generated by a box-conditioned ResNet34 U-Net trained on the open-access TN3K dataset. The fusion head was selected via in-training validation by adding a limited residual correction to the frozen image logits, and the final model was locked before testing. The three-seed ensemble achieved an AUC of 0.949, a PR-AUC of 0.978, an accuracy of 0.890, a balanced accuracy of 0.867, and a Macro-F1 score of 0.862 on 1,000 test samples. These test results were supported by 95% confidence intervals obtained from 10,000 class-stratified bootstrap repetitions. Compared with the same image-only backbone aggregated identically as a three-seed ensemble, the proposed model improved AUC by 0.019. These findings show that fusing multiscale CNN representations with measurable shape, boundary, and echotexture information improves upon the image-only baseline and brings clinically interpretable features into the model. As the evaluation was confined to a single open-access dataset without external validation, the reported performance should be regarded as methodological evidence rather than a demonstration of clinical readiness.