The study developed an improved YOLOv8 average aggregate downsampling (ADown) model for the detection and classification of A1–A3 intertrochanteric femoral fractures according to the AO/OTA classification in radiographs. They used a retrospective design with 976 radiographs from hospital archives that were processed, annotated by orthopedists, and divided into training and test sets. The model replaced traditional convolutional downsampling modules with ADown for better extraction of small fracture features and used data augmentation techniques. Detection accuracy increased by 7.3% for type A1, 3.5% for A2 and 7.8% for A3. The number of model parameters decreased by 12.3% and computational complexity (FLOP) by 9.8%, enabling deployment on edge devices. The YOLOv8-ADown model provides an effective solution for fracture detection and clinical diagnosis support. Future work will focus on improving data collection and multicenter validation.