# MFSD-YOLO: A multi-scale feature detection network for pediatric wrist abnormalities in radiographic images

## TL;DR

MFSD-YOLO is a new AI model that improves detection of wrist fractures in children using multi-scale feature analysis for better accuracy and speed.

## Contribution

Introduces MFSD-YOLO, a novel multi-scale detection model with modules like CSP_PMSFA, FPSConv, and C2BRA for pediatric wrist abnormality detection.

## Key findings

- MFSD-YOLO achieves 69.7% mAP@0.5, a 5.3% improvement over YOLOv11 on the GRAZPEDWRI-DX dataset.
- The model maintains 10.8M parameters and 3.2 ms inference speed, making it suitable for real-world clinical use.
- Modules like SlideLoss help address class imbalance and improve rare target detection.

## Abstract

Timely and accurate diagnosis of wrist abnormalities, especially distal radius and ulna fractures, is critical in children and adolescents, yet complicated by skeletal immaturity, overlapping anatomy, and low-contrast fracture lines. To address these challenges, we propose MFSD-YOLO, a multi-scale detection model for pediatric wrist abnormality analysis. The model integrates a Cross-Stage Partial Progressive Multi-Scale Feature Aggregation (CSP_PMSFA) module inspired by GhostNet that applies lightweight multi-scale convolutions on partial channels with partial convolution and residual connections to reduce redundancy and enhance shallow texture and subtle fracture sensitivity. The Feature Pyramid Shared Convolution (FPSConv) module replaces pooling with shared dilated convolutions to expand the receptive field and capture multi-scale context without added cost. The C2 Bi-Level Routing Attention (C2BRA) module, based on C2PSA, uses regional routing and local enhancement to refine focus on relevant areas while balancing accuracy and speed. The Recursive Gradient Dynamic Feature Pyramid Network (RepGDFPN) optimizes top-down and bottom-up multi-scale fusion, reducing semantic loss and improving robustness. Finally, the Sliding Weight Adaptive Loss (SlideLoss) addresses class imbalance, enhancing detection of rare targets. Evaluated on the GRAZPEDWRI-DX dataset, MFSD-YOLO achieves 69.7% mAP@0.5, representing a 5.3% improvement over the baseline YOLOv11, while maintaining 10.8M parameters and 3.2 ms inference speed. These results validate the model’s effectiveness and its potential for real-world deployment in clinical pediatric radiographic analysis.