Generative Artificial Intelligence for Recognition of Surgical Site Complications: the PRISCA Project

Abstract

The project PRISCA develops digital tools for postoperative monitoring and early detection of wound complications
(e.g., infections) through telemedicine solutions. The approach addresses concrete clinical needs, including the follow-up of
patients discharged from centers of surgical excellence located far from home, the optimization of hospital access, and the
early identification of at-risk situations. The output of the project will be a telemonitoring platform (mobile and web apps),
artificial intelligence modules for wound image analysis, and informative content for patients. The technological architecture
is designed to support timely intervention in case of wound-related complications, while also reducing unnecessary in-person
visits. The project’s main innovation is the use of an Artificial Intelligence module. It aims to enable healthcare professionals
to perform automated analysis of wound images for early detection of post-surgical complications. The limited availability
of public data was tackled by applying a data augmentation method and by integrating a generative AI model. It creates new
synthetic images based on real images and textual prompts. All generated images had been validated by clinicians and then
included in the final dataset. This approach ensures the model learns from a diverse set of images, increasing its robustness
and accuracy. The adopted detection model is YOLOv11, which localizes the wound and performs a pathological/nonpathological
classification. Results show good localization and promising classification accuracy. We performed a comparison
between the model trained on the original dataset and the version enhanced with synthetic data, in order to assess relative
improvements. These comparisons will help refine the model for better performance in real-world scenarios. The first results
show an increase in the performance of the model with augmented data but more systematic comparisons are needed.
Additional real images from a proprietary dataset currently being collected will also be integrated, further enhancing the AI's
ability to identify early complications.