<p>This study aims to address the issues of missed and misdiagnosed supernumerary teeth during the mixed dentition stage in children by investigating how computer-aided diagnosis can improve the detection rate of supernumerary teeth during routine pediatric dental imaging examinations. By integrating the clinical diagnostic expertise of professional dentists regarding supernumerary teeth, the study proposes a dual-collaborative classification method for supernumerary teeth in children, which is achieved by establishing two collaborative mechanisms: “localization–classification” and “global–local.” Drawing on the clinical approach of “locate first, then judge,” this method uses a localization module to precisely identify regions of interest (ROIs). It employs ResNeXt as the backbone network, incorporates a feature-fusion spatial adaptive attention module, and utilizes Efficient-KAN as the classification head to enhance the nonlinear representation of complex dental features in children, thereby accurately capturing key regional characteristics. Inspired by the way doctors confirm diagnoses by integrating overall images with local details, the method fuses category probability information to integrate results from both global and local branches, achieving precise classification. Experiments demonstrate that this method performs exceptionally well on the CSTD dataset for classification tasks, achieving an accuracy of 0.99, a precision of 0.94, a recall of 0.96, an F1 score of 0.95, and a Kappa coefficient of 0.94. All of these classification metrics outperform existing methods, enabling efficient differentiation between normal teeth and supernumerary teeth while reducing both misdiagnosis and missed diagnosis rates. This approach provides a viable solution for early detection and holds clinical application value.</p>