This study addresses the clinical challenge of scarce annotated data in cervical cancer histopathological image classification by proposing a multi-task collaborative few-shot learning method based on a Siamese network architecture. Through the construction of dual-path feature embedding spaces and dynamic multi-task matching strategies, the model achieves precise benign-malignant classification of pathological images under single-sample conditions (1-shot learning). The innovative integration of channel-spatial dual attention mechanisms and enhanced multi-task design yields 1-shot classification accuracies of 90.4% and 91.3% on the CAISHI Digital Biobank (DiB) and food11 datasets, respectively, representing a 13.6% improvement over conventional methods. Experiments demonstrate that the pathology-preserving data augmentation strategy effectively mitigates domain shift issues, while the dynamic weight multi-task optimization framework achieves near-perfect classification performance (close to 100%) in 5-shot scenarios. This method provides an interpretable, low-cost AI-assisted diagnostic solution for primary healthcare institutions, demonstrating significant application value in early cervical cancer screening.