LightRefine-PCXR: A Lightweight Refinement Framework for Efficient Medical Device Suppression in Pediatric Chest X-Rays

Mingze Jiang, Xueyang Li, John Kheir, Alec Girten, Yiyu Shi

Keywords: Medical Device Suppression, Pediatric Chest X-ray, Resource-efficient Refinement, Plug-in Module, Generative Inpainting

TL;DR: LightRefine-PCXR is a lightweight, anatomy-aware refinement framework for suppressing medical devices in pediatric chest X-rays, providing high-quality restoration with low training cost and limited pediatric data for clinical deployment.

Abstract: In pediatric chest radiography, indwelling support devices (e.g., tubes and lines) are ubiquitous and often obscure critical thoracic structures, complicating radiologic interpretation and reducing the reliability of automated analysis methods. Although generative inpainting has advanced rapidly, reliable deployment in pediatric chest radiographs remains challenging. Subtle anatomical cues must be preserved under substantial domain shift, while full adaptation of large backbones is often impractical because of limited pediatric data and constrained clinical compute budgets. To address these limitations, we propose LightRefine-PCXR, a lightweight, backbone-agnostic refinement framework for suppressing medical devices in pediatric chest X-rays (PCXRs). LightRefine-PCXR follows a two-stage strategy: a frozen pretrained inpainting backbone first produces a coarse device-removed estimate, and a compact anatomy-aware refiner then predicts mask-constrained residual corrections to restore local structures while preserving all unmasked pixels exactly. This plug-in design substantially reduces trainable parameters and peak GPU memory compared with end-to-end fine-tuning, yet consistently improves reconstruction fidelity and perceptual quality across diverse inpainting paradigms, including CNN-, transformer-, and diffusion-based models. Comprehensive in-domain and cross-dataset experiments demonstrate robust device suppression and strong generalization in low-data pediatric settings, highlighting the practicality of LightRefine-PCXR for real-world pediatric radiology workflows. Our model will be publicly available upon acceptance.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 124