Automated classification of osteoporotic vertebral compression fractures in DR images based on improved YOLO26n-seg

Osteoporotic vertebral compression fractures (OVCFs) are prevalent fragility fractures among the middle-aged and elderly populations. The widely used manual Genant classification for OVCF assessment is limited by low efficiency, high subjectivity, and a significant rate of underreporting on imaging. Furthermore, dedicated deep learning models for accurate classification remain scarce.

To address these clinical challenges, we proposed an improved YOLO26n-seg-SAConv model by substituting the traditional convolutional downsampling module of YOLO26n-seg with a Switchable Atrous Convolution (SAConv) module. A retrospective dataset comprising 483 digital radiography (DR) images of OVCFs was collected from the First Affiliated Hospital of Soochow University and its affiliated institutions between January 2021 and January 2026. After preprocessing, the data were divided into training and test sets at an 8:2 ratio and augmented with Gaussian noise.

The improved model achieved a mean average precision at an Intersection over Union of 0.5 (mAP50) of 74.2%, representing a 3.9% improvement over the original model. Concurrently, the computational complexity (floating-point operations per second, FLOPs) was reduced by 16.67%. Notably, the modified model demonstrated significantly enhanced classification performance for Type 1 and Type 3 fractures, which conventionally suffer from smaller sample sizes.

The proposed lightweight YOLO26n-seg-SAConv model is suitable for edge deployment. This methodological study verifies the feasibility of automated Genant classification of OVCFs on DR images, laying a technical basis for its further development into an auxiliary triage tool in subsequent clinical translation research.