SurgRFO: Foundation Model Based Compositional Synthesis of Critical Retained Foreign Objects in Intraoperative Chest X-rays

Yuanyun Hu †, Yuli Wang †, Noemi Acevedo Rodriguez, Ronald Yang, Wen-Chi Hsu, Siwei Luo, Zihao Bai, Jing Wu, Yuwei Dai, Shaoju Wu, Jonathon Lindquist, Justin Honce, Premal Trivedi, Zhicheng Jiao, Ihab Kamel, Elliott Haut, Pamela Johnson, John Eng, Cheng Ting Lin, Nan Su, Bo Chen *, Sun Yu, Harrison Bai * · 2026

Read SurgRFO: Foundation Model Based Compositional Synthesis of Critical Retained Foreign Objects in Intraoperative Chest X-rays by Yuanyun Hu †, Yuli Wang †, Noemi Acevedo Rodriguez, Ronald Yang, Wen-Chi Hsu, Siwei Luo, Zihao Bai, Jing Wu, Yuwei Dai, Shaoju Wu, Jonathon Lindquist, Justin Honce, Premal Trivedi, Zhicheng Jiao, Ihab Kamel, Elliott Haut, Pamela Johnson, John Eng, Cheng Ting Lin, Nan Su, Bo Chen *, Sun Yu, Harrison Bai * (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-24787.txt.

Introduction

Yuanyun Hu and others introduce SurgRFO, a two-stage framework for retained foreign objects in chest X-rays. Such objects are rare but high-risk, and scarcity limits training. Stage one fine-tunes a Roentgen model on surgical images. It generates object-free backgrounds with anatomy, lines, tubes, and traits. Stage two trains a light generator on object patches from few cases. It composites diverse instances onto backgrounds with Poisson fusion. Blinded clinician ratings match real-image realism. Synthesized data improve sensitivity at low false-positive rates.

Catalog record

CREATORYuanyun Hu †, Yuli Wang †, Noemi Acevedo Rodriguez, Ronald Yang, Wen-Chi Hsu, Siwei Luo, Zihao Bai, Jing Wu, Yuwei Dai, Shaoju Wu, Jonathon Lindquist, Justin Honce, Premal Trivedi, Zhicheng Jiao, Ihab Kamel, Elliott Haut, Pamela Johnson, John Eng, Cheng Ting Lin, Nan Su, Bo Chen *, Sun Yu, Harrison Bai *
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21531
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.24787

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SurgRFO: Foundation Model Based Compositional Synthesis of Critical Retained Foreign Objects in Intraoperative Chest X-rays Yuanyun Hu †, Yuli Wang †, Noemi Acevedo Rodriguez, Ronald Yang, Wen-Chi Hsu, Siwei Luo, Zihao Bai, Jing Wu, Yuwei Dai, Shaoju Wu, Jonathon Lindquist, Justin Honce, Premal Trivedi, Zhicheng Jiao, Ihab Kamel, Elliott Haut, Pamela Johnson, John Eng, Cheng Ting Lin, Nan Su, Bo Chen *, Sun Yu, Harrison Bai * Abstract Critical retained foreign objects (RFOs) on intraoperative chest radiographs are rare but high-risk events. Their scarcity limits robust automated detection model training and generalization. We introduce SurgRFO , a two-stage synthesis framework for generating realistic RFO-present intraoperative chest X-rays. In Stage 1 , a Roentgen chest X-ray foundation model is fine-tuned on surgical-domain images to generate realistic RFO-free backgrounds that preserve anatomy, indwelling lines and tubes, and intraoperative imaging characteristics. In Stage 2 , a lightweight generator trained on localized RFO patches from limited positive cases synthesizes diverse RFO instances, which are composited onto generated backgrounds using conditional Poisson fusion to improve photometric consistency. We evaluate SurgRFO through (i) a blinded clinician study assessing realism and clinical plausibility, and (ii) downstream detection experiments in which synthesized data are used to augment Faster R-CNN, YOLOv8, and RetinaNet. SurgRFO consistently improves sensitivity at low false-positive-per-image (FPPI) operating points on internal and external test sets. Clinician ratings indicate that the synthesized images achieve realism comparable to real intraoperative images. Ablation analyses further examine fusion strategies and synthesis scale. Ethical safeguards for synthetic surgical data are also discussed. The code is publicly available on GitHub at https://anonymous.4open.science/r/SurgRFO-7CB5/ . † † footnotetext: These two authors contributed equally to this work. * * footnotetext: Corresponding author: [email protected] and [email protected] 1 Introduction Critical retained foreign objects (RFOs), such as surgical sponges and needle fragments, are rare but high-impact patient safety events. Unlike non-critical, intentionally placed objects, critical RFOs are unintended remnants that may cause severe complications, re-operations, and substantial medico-legal and financial burden. [ 18 , 8 , 17 , 13 , 16 , 22 , 23 ] . Clinically, intraoperative imaging aims to rule out RFOs before procedure completion, where missed detections can have serious consequences. However, work explicitly targeting critical RFOs remains limited. Most prior studies focus on standardized chest X-rays (CXRs) and predominantly address non-critical objects [ 30 , 24 , 21 , 10 ] . Robust detection of critical RFOs in real-world intraoperative X-rays, therefore, remains an open and clinically significant challenge. Critical RFO detection relies on intraoperative radiographs acquired under far less controlled conditions than standard CXRs, with cluttered scenes, overlapping instruments, variable anatomy exposure, and heterogeneous imaging settings. Two major challenges arise. First, visual cues are often subtle and confounded by surgical context. Sponges or needle fragments may be blood-soaked, partially occluded, or blended with surrounding tissue, producing weak signals that are difficult to recognize, even for experienced clinicians. [ 11 , 14 ] . False negatives are therefore not uncommon, and prior studies suggest that intraoperative imaging may fail to detect a substantial fraction of RFO cases [ 2 ] . Second, positive cases are extremely rare and difficult to scale. Even at a large hospital, only low hundreds of cases may accumulate over decades, and no public dataset dedicated to critical RFOs exists. Together, these constraints create a fundamental limitation: accurate detection requires sensitivity to subtle local patterns, yet available data provide limited exposure to such events, challenging conventional supervised learning approaches. Figure 1 : Overview of the Intraoperative RFO Bench and SurgRFO pipeline. (a) Multi-institutional intraoperative radiographs are curated and expert-annotated as real positive (critical RFO) and negative cases. (b) SurgRFO applies a two-stage synthesis framework, combining surgical X-ray background generation with localized RFO sampling and fusion to create annotated synthetic images. (c) Synthetic data are used for training augmentation and evaluated on held-out real cases for critical RFO detection and classification. (d) A radiologist-led Turing test assesses visual realism through confidence-based quality ratings and image discrimination tasks. Synthetic data augmentation is a natural strategy to mitigate data scarcity. Recent advances in diffusion models (DMs), which generate images through iterative denoising of a learned reverse process [ 5 ] , have enabled high-fidelity and diverse X-ray synthesis [ 1 , 27 , 29 , 25 , 31 ] , supporting augmentation for imbalanced medical imaging tasks [ 9 , 15 , 7 , 12 , 6 , 28 ] . However, fully end-to-end generation of critical RFO radiographs remains unreliable. The same rarity and subtle visual signatures that hinder detection also limit generative fidelity: standard pipelines struggle to simultaneously preserve complex surgical context and localized RFO cues, often leading to visually implausible samples for training detectors (see comparisons in Section 3 ). To address this gap, we propose SurgRFO , a two-stage synthesis framework for generating surgical radiographs with critical RFOs. Stage 1 learns a strong prior over cluttered surgical radiographs by training a latent diffusion model (LDM) as a generic surgical X-ray generator. Stage 2 trains a lightweight U-Net to model local critical RFO appearance, then conditionally fuses the sampled RFO content into Stage 1 backgrounds via a Poisson-based blending mechanism guided by a prior over plausible RFO locations and sizes. We also release Intraoperative RFO Bench , the first publicly available dataset dedicated to critical RFOs, comprising 144 expert-annotated positive cases curated from 18 years of surgical radiographs from two health systems (Site 1 and Site 2). Augmenting training with SurgRFO yields consistent improvements across multiple detector families. In a radiologist-led Turing-style evaluation, synthetic images are frequently indistinguishable from real cases. Together, these results demonstrate the value of releasing this benchmark and the effectiveness of the proposed framework for advancing critical RFO detection. 2 Method 2.1 Dataset Curation Under institutional review board approval, we retrospectively curated surgical chest radiographs from the Site 1 Health System, comprising 144 critical RFO-positive cases and 944 negative cases with no RFOs. An additional 20 critical RFO-positive cases were collected from Site 2 for external evaluation. Positives were identified via structured report search and expert verification using radiology reports and operative notes. All images were de-identified and annotated by trained radiologists under a standardized protocol, including image-level labels and object-level annotations (bounding boxes or polygons), with each foreign object categorized as critical or non-critical. Patient-level splits into training, validation, and test sets were applied consistently across cohorts to prevent information leakage. Training images were preprocessed with center cropping, resolution normalization, and artifact filtering. 2.2 SurgRFO Let \Omega\subset\mathbb{R}^{2} denote the image lattice, and let I\in\mathbb{R}^{H\times W} denote a radiograph defined on \Omega . Two datasets are considered, \mathcal{D}_{\mathrm{bg}}=\{I_{i}\}_{i=1}^{N_{\mathrm{bg}}} and \mathcal{D}_{\mathrm{rfo}}=\{(I_{j},M_{j},y_{j})\}_{j=1}^{N_{\mathrm{rfo}}} . The set \mathcal{D}_{\mathrm{bg}} contains surgical-domain negative radiographs that provide RFO-free backgrounds. The set \mathcal{D}_{\mathrm{rfo}} contains radiographs with annotated critical RFOs. The mask M_{j}\subset\Omega specifies the RFO support, and the label y_{j}\in\mathcal{Y} specifies the semantic type. SurgRFO is a two-stage synthesis framework that decouples global background modeling from local critical RFO modeling. Stage 1 learns a generic prior over surgical radiograph backgrounds. Stage 2 learns a local sampler for critical RFO appearance and composes sampled RFO evidence onto Stage 1 backgrounds through a controlled fusion mechanism. Stage 1: Generic Surgical Radiograph Generator. A background generator is defined as a conditional LDM f_{\theta} parameterized by

Read the complete text