：Cytologic examination of thyroid fine-needle aspiration biopsy whole-slide image (FNAB-WSI) is crucial for the diagnosis of papillary thyroid carcinoma or benign nodular hyperplasia. Due to the ultra-high resolution in sample-level FNAB-WSI, sample-level classification using deep networks consumes computational resources of considerable scale. Given that the sample-level FNAB-WSI has both global and cell cluster local detail features, a lightweight sample-level FNAB-WSI classification method with global-local feature fusion is proposed. Firstly, the global features are extracted using lightweight GhostNet, the feature map size is controlled by setting the convolutional stride, and the local features are obtained by feature slicing and fusion. Then, the global and local features are fused into global-local features after max-pooling and dimensionality reduction, respectively. Finally, the global-local features are fully connected to classify the benign-malignant FNAB-WSI by the softmax classifier. On the self-build FNAB-WSI sample-level dataset, our method surpasses other lightweight methods in all performance indicators, with 89.9% precision, 91.2% recall, 91.7% Acc, and 92.5% AUC, respectively, while the number of parameters is comparable to 6.1M, demonstrating a tradeoff result. The proposed method not only improves the accuracy of sample-level classification, but also optimizes the computational efficiency of the model by reducing the number of parameters, providing an effective auxiliary tool for clinical diagnosis of thyroid cancer.