[Paper Review] Attention-based Pin Site Image Classification in Orthopaedic Patients with External Fixators

Yubo Wang, Marie Fridberg+5 more|arXiv (Cornell University)|Mar 25, 2026

Diabetic Foot Ulcer Assessment and Management0 citations

TL;DR

The paper presents an attention-based CNN with an Efficient Redundant Reconstruction Convolution (ERRC) to classify pin site images into inflamed/infected vs. non-infected, achieving high AUC with a small model on a small, manually annotated dataset.

ABSTRACT

Pin sites represent the interface where a metal pin or wire from the external environment passes through the skin into the internal environment of the limb. These pins or wires connect an external fixator to the bone to stabilize the bone segments in a patient with trauma or deformity. Because these pin sites represent an opportunity for external skin flora to enter the internal environment of the limb, infections of the pin site are common. These pin site infections are painful, annoying, and cause increased morbidity to the patients. Improving the identification and management of pin site infections would greatly enhance the patient experience when external fixators are used. For this, this paper collects and produces a dataset on pin sites wound infections and proposes a deep learning (DL) method to classify pin sites images based on their appearance: Group A displayed signs of inflammation or infection, while Group B showed no evident complications. Unlike studies that primarily focus on open wounds, our research includes potential interventions at the metal pin/skin interface. Our attention-based deep learning model addresses this complexity by emphasizing relevant regions and minimizing distractions from the pins. Moreover, we introduce an Efficient Redundant Reconstruction Convolution (ERRC) method to enhance the richness of feature maps while reducing the number of parameters. Our model outperforms baseline methods with an AUC of 0.975 and an F1-score of 0.927, requiring only 5.77 M parameters. These results highlight the potential of DL in differentiating pin sites only based on visual signs of infection, aligning with healthcare professional assessments, while further validation with more data remains essential.

Motivation & Objective

• Collect and label a pin-site image dataset (NCH) for infection vs. non-infection based on visual appearance.
• Develop an attention-based CNN architecture robust to pin interference at the skin-pin interface.
• Introduce Efficient Redundant Reconstruction Convolution to enrich features while reducing parameters.
• Evaluate model performance against baselines and provide model interpretability with Grad-CAM.
• Discuss limitations and potential for remote, non-contact wound assessment in orthopaedics.

Proposed method

• Construct a dataset of 666 pin-site images from Nationwide Children’s Hospital and detect pin locations with YOLOv5 to define ROIs.
• Propose a five-block CNN with CBAM attention and ERRC blocks to enrich feature maps and suppress pin interference.
• Address class imbalance with Focal loss (alpha=0.15, gamma=2).
• Train without pretraining on a 70/30 train/test split (with 80/20 train/val inside the training set) and report Precision, Recall, AUC, and F1.
• Evaluate parameter efficiency and inference speed; compare against VGG-16/19, ResNet-50, EfficientNetV2-b0, with and without focal loss and ERRC.

Experimental results

Research questions

• RQ1 Can a lightweight, attention-based CNN accurately classify pin-site images into inflamed/infected vs. non-infected based on visual appearance alone?
• RQ2 Does the ERRC module reduce parameter count while enriching feature maps to improve infection detection performance?
• RQ3 How does focal loss affect model performance on an imbalanced pin-site dataset?
• RQ4 How does the proposed model compare to standard architectures on this small medical image dataset?
• RQ5 Is the model’s attention localized to meaningful skin areas around pins, as shown by Grad-CAM?

Key findings

• The proposed ERRC model achieves an AUC of 0.975 and F1 of 0.927 with 5.77M parameters.
• Compared baselines include VGG-16 (Prec 99.4, Recall 53.2, AUC 0.807, F1 0.69), VGG-19 (Prec 98, Recall 52.69, AUC 0.822, F1 0.685), ResNet-50 (Prec 70, Recall 80.46, AUC 0.931, F1 0.749), EfficientNetV2-b0 (Prec 90, Recall 88.23, AUC 0.954, F1 0.889).
• Model with FL (Focal Loss) yields Prec 89.1, Recall 90, AUC 0.957, F1 0.895; model without FL yields Prec 95.4, Recall 79, AUC 0.907, F1 0.864.
• ERRC reduces parameters by 30.6% vs inverted residual block and outperforms in AUC and F1.
• Inference time: ERRC model achieves mean 2.37 ms, 42.4 FPS (batch size 1).
• Grad-CAM visualizations show the model focusing on wound skin near pins, mitigating pin interference.