To address the challenges of strong background interference，difficulty in recognizing multi-scale lesions，and the need for lightweight deployment in apple disease detection within complex orchard environments，an improved lightweight apple disease detection model named EGA-DETR was proposed based on RT-DETR. The model was enhanced in three key aspects. First，an RGLAN lightweight feature aggregation module was designed to reduce redundant computations and enhance feature representation through feature splitting and re-parameterized convolution. Second，a BiFPN-GLSA multi-scale feature fusion module was developed to improve the representation of lesions at various scales via bidirectional feature transmission and a global–local self-attention mechanism. Third，an Inner-Shape-IoU loss function was introduced to improve model’s localization accuracy for irregularly shaped lesion targets. Experimental results showed that EGA-DETR achieved 91.8% precision，88.9% recall，and 92.1% mAP₅₀ on the apple disease dataset，with improvements of 3.5，3.5，and 1.5 percentage points over the baseline RT-DETR-18，respectively. Meanwhile，the number of model parameters was reduced to 11.8×10⁶，which is 40.4% fewer than that of the baseline model，and the inference speed reached 120 frames per second. In summary，the EGA-DETR model achieved a favorable balance between detection accuracy and computational efficiency，providing robust technical support for accurate and real-time apple disease detection.