With the ongoing advancement of digitalization and mobile internet, the spread of fake news on social media has become increasingly pervasive. However, existing approaches often fail to effectively model the interactions between comments and multimodal tweet content, primarily focusing on single-dimensional comment filtering and neglecting other critical factors in refuting false information. In this study, we propose a multimodal comment self-selection model (CSSFND) for fake news detection. This model identifies comments that effectively support or refute the veracity of a tweet based on true and false attention scores, thereby providing valid support for true information and strong rebuttals for false claims. Unlike fake news detection methods that rely solely on pre-trained models, our approach integrates the collective intelligence of users while leveraging the internal knowledge of these models. We conduct comprehensive comparative experiments on two real-world social media datasets and validate the effectiveness of the comment self-selection module through ablation studies, achieving near state-of-the-art results on both datasets.