Urban intersections, critical hotspots for severe traffic crashes, demand advanced analytical approaches, particularly in developing nations where such methods remain underexplored. This study introduces a novel data mining framework integrating deep learning to analyze collision patterns at urban intersections in Qazvin, Iran. Utilizing 245 crash records (2021–2023), the research applied k-means clustering to categorize variables, regression trees for classification, and a deep convolutional neural network (DCNN) to evaluate data clusters. Three primary crash-influencing factors emerged: vehicle type, human factors, and lighting conditions. Training the DCNN on 2021–2022 data and testing it on 2023 data yielded 97% accuracy in predicting crash determinants. Findings highlighted unique crash characteristics across intersections, emphasizing context-specific risk factors. The model enables precise identification of variable importance, offering actionable insights for targeted safety interventions. This approach bridges methodological gaps in crash analysis for developing regions, demonstrating the efficacy of hybrid data mining and deep learning in enhancing intersection safety planning. By prioritizing key risk variables and enabling predictive analytics, the framework supports data-driven policymaking to mitigate crash severity and frequency in urban settings.