SurgRFO uses foundation models to synthesize X-rays of retained surgical objects

Medical AI researchers have developed SurgRFO, a novel two-stage synthesis framework that generates realistic intraoperative chest X-rays containing critical retained foreign objects (RFOs) — surgical tools or materials accidentally left inside patients during surgery. The scarcity of real positive cases makes training automated detection models extremely difficult. SurgRFO addresses this by first fine-tuning a Roentgen chest X-ray foundation model on surgical-domain images to produce high-fidelity RFO-free backgrounds that accurately preserve anatomy, indwelling lines, tubes, and intraoperative imaging characteristics. In the second stage, a lightweight generator trained on localized RFO patches from limited positive cases synthesizes diverse object instances, which are then composited onto the generated backgrounds using conditional Poisson fusion for photometric consistency.

The team validated SurgRFO through a blinded clinician study where surgeons rated the synthetic images as having realism comparable to real intraoperative radiographs. Downstream detection experiments using Faster R-CNN, YOLOv8, and RetinaNet showed that augmenting training data with SurgRFO-generated images consistently improved sensitivity at low false-positive-per-image (FPPI) operating points on both internal and external test sets. Ablation studies examined different fusion strategies and synthesis scales. The paper also discusses ethical safeguards for synthetic surgical data. This approach could significantly bolster patient safety by enabling robust AI detection of RFOs, a rare but life-threatening complication.