Accurate detection of apple diseases is of great significancecrucial for ensuring fruit quality and sustainable industrial development. However, in complex orchard environments, existing detection methods face challenges such as difficulty in multi-scale target recognition, severe background clutter interference, and limited computational resources in complex orchard environments. This study proposes EGA-DETR, a lightweight apple disease detection network based on an improved RT-DETR architecture, which achieves an optimal balance between detection accuracy and computational efficiency through innovative integration of efficient feature extraction, multi-scale feature fusion, and geometry-aware optimization strategies. Theis research develops"s technical innovations are threefold across three dimensions: Firstly, an lightweight RGLAN lightweight feature aggregation module is designed, which employs a feature splitting strategy and re-parameterized convolution mechanism to perform deep transformations on only partial input features, thereby enhancing feature representation capability while reducing computational complexity. Secondly, a BiFPN-GLSA module is constructed, which integrates Bidirectional Feature Pyramid Network with Global-Local Self-Attention mechanism to achieve comprehensive interaction and adaptive focusing of features across different scales. ThirdFinally, an Inner-Shape-IoU loss function is introduced, which improves localization accuracy for irregular lesions by establishing structured representations of geometric shapes within bounding boxes. Experimental results demonstrate that EGA-DETR achieves a detection precision of 92.4%, recall of 89.6%, and mAP50 of 92.6%, representing improvements of 3.5, 2.9, and 1.4 percentage points respectively compared to the baseline model. Meanwhile, the model contains only 11.8M parameters, a 40.4% reduction from the baseline, while maintaining real-time detection speed of 120FPS. Comparative experiments with mainstream algorithms such as YOLOv11 and YOLOv8 further validate the comprehensive advantages of this method in terms of detection performance and computational efficiency. This research provides ana practicaleffective technical solution for real-time disease monitoring in agricultural scenarios and significant application valuedemonstrates significant potential for advancing smart agriculture development.