Cucumber leaf diseases seriously impact crop yield and quality, demanding cost-effective, efficient diagnostic tools. Currently, field inspections by experts are primary but costly, inefficient, and lack scalability. Moreover, the requirement of significant computing resources and the presence of complex backgrounds in agricultural settings both constrain the applicability of most existing leaf disease detection algorithms. To address these limitations, this paper proposes a Deformable Object Detection Network (DODN) for lightweight cucumber leaf disease detection in natural scenes. By incorporating deformable convolution and Transformer components into the YOLOv5 framework, DODN can accurately identify and localize cucumber diseases, even in the presence of irregular target shapes and complex backgrounds. Rigorous experiments conducted on a comprehensive cucumber disease dataset comprising real-world, complex field images as well as publicly available datasets demonstrate that DODN achieves state-of-the-art detection accuracy, with a mean average precision (mAP) of 67.2%. Furthermore, its compact model size of 3.7 MB and computational efficiency of 3.9 GFLOPs make it a hardware-friendly choice for lightweight and portable cucumber disease detection devices.