Related Experiment Videos
Detecting Incisional Surgical Site Infections on Wound Images Through Deep Learning
Nathan Bontekoning1,2, Hiske Huisman1,2, Pedro J Segura Cabrera3
1Department of Surgery, Amsterdam, Amsterdam UMC location University of Amsterdam, the Netherlands.
Importance:
Surgical site infections (SSIs) are common postoperative complications, often detected after patients leave the hospital, especially in times of modern fast-track recovery protocols. At-home monitoring through telemedicine can shorten delay to diagnosis. Manual wound image review in telemedicine is time-consuming, shows interobserver variability, and may miss early signs.
Objective:
To develop and externally validate a deep learning model to detect wounds suggestive of SSIs from wound photographs.
Design, Setting, And Participants:
This diagnostic/prognostic study involved the development and evaluation of an artificial intelligence model. Surgical wound images from patients in existing databases from multiple specialties were used for model development and internal testing. Prospectively collected images from multiple specialties in a large academic center were used for external validation.
Exposures:
Each image was labeled by multiple physician as either suggestive or nonsuggestive for SSI. A convolutional neural network based on InceptionV3 was developed using transfer learning. The retrospective dataset was split into training (70%), validation (15%), and internal test (15%) sets.
Main Outcomes And Measures:
Model performance was evaluated by area under the receiver operating characteristic (AUC) curve, calibration curves, and clinical utility analyses, both in the internal test set and the independent external validation dataset. Gradient-weighted class activation mapping heat map visualizations were added to show focus on wound regions in an image.
Results:
The model was trained on 4978 wound images and externally validated on 407 images from 95 patients. In the internal validation set, the model showed strong discrimination with an AUC of 0.91 (95% CI, 0.88-0.93), with good calibration. In the external validation set, the model achieved an AUC of 0.82 (0.75-0.90). Decision curve analyses showed a net benefit of the model.
Conclusions And Relevance:
This study developed and validated a deep learning model for detecting SSIs from wound images, showing high potential for automated triage of cases if clinically integrated.