To achieve rapid and precise diagnosis of citrus fruit surface diseases in complex orchard environments, this paper presents YOLO-CitrusDet, an application-oriented detection model based on YOLOv12 [5]. To address challenges such as minute disease targets, highly similar visual characteristics among multiple diseases, and irregular lesion shapes, this study integrates three task-adapted components rather. First, a Citrus Disease Attention Enhancement (CDAE) module, adapted from SCSA/CBAM-style spatial-channel attention, is integrated to strengthen the model's ability to distinguish subtle symptoms and suppress complex background interference. Second, a Citrus-oriented Multi-scale Feature Fusion Neck (CMFF-Neck), inspired by the attention scale sequence fusion idea of ASF-YOLO [3], is adopted to adaptively integrate features across different levels and enhance tiny-lesion detection. Finally, a Lesion Shape-aware Regression Loss (LSR-Loss), adapted from Shape-IoU [4], is used to improve bounding box fitting for irregular lesions and alleviate the optimization bias caused by the imbalance between "healthy" and "diseased" samples in the dataset. Experiments conducted on a self-built dataset containing five citrus fruit surface states (Huanglongbing, melanose, black spot, canker, and healthy) demonstrate that the proposed method achieves 90.7% in [email protected], which is a 2.8% improvement over the baseline YOLOv12 while maintaining efficient detection speed. This provides an effective application solution for intelligent monitoring of citrus fruit surface diseases.