<p>Utilizing X-ray screening for prohibited item inspection is crucial for preventing crime and maintaining public safety. Since obtaining X-ray images that contain prohibited items is challenging, few-shot learning has become a prominent approach for this. However, unlike natural images with rich features, most existing few-shot segmentation methods cannot be directly applied to the X-ray images suffering from overlap, complex background, and size variation. To counter these challenges, we present the novel ideologies of target category focus and self-adaptive prototype guidance to improve the automatic inspection of prohibited items. Specifically, for target category focus, we leverage the cross-attention mechanism to leverage support masks for guiding information interaction between support and query features, thereby enhancing target class feature regions while suppressing background class feature regions. For self-adaptive prototype guidance, we develop an algorithm of generating superpixel prototypes to mitigate the loss of spatial detail that has been caused by mask average pooling. Additionally, we design an extended angular similarity assessment method to comprehensively evaluate the angular and magnitude similarities between query features and support prototypes, so as to suppress the complex background. Finally, a prototype expansion match is performed to ensure alignment between prototypes and semantic information, thereby providing more precise guidance. To validate the effectiveness of our proposed method, we adapt the existing PIXray dataset based on PASCAL-5i. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches on the adapted PIXray dataset.</p>