ReFINE-Lung: REcalibration and False positive INference Engine for Consensus-Based Lung Nodule Detection

Center for Virtual Imaging Trials, Carl E. Ravin Advanced Imaging Laboratories, Department of Radiology,Duke University School of Medicine, Durham, NC, 27708, USA

Abstract

The clinical utility of AI models trained on simulated or publicly annotated datasets hinges on their ability to generalize to large-scale, real-world imaging data—an enduring challenge due to the high cost and complexity of expert annotations. To address this, we introduce ReFINE-Lung, a modular, consensus-driven framework for curating high-quality pseudo-labels from unannotated clinical lung CT scans. ReFINE-Lung begins with ensemble-based nodule candidate generation using diverse AI models trained on simulated, public, and institutional data. It then refines predictions through a four-stage ReFINE Block: (1) affine calibration of model outputs, (2) unsupervised prediction alignment to the clinical data distribution, (3) disease prevalence estimation via kernel density modeling, and (4) cost-sensitive thresholding. By combining multi-model agreement with morphological filters and statistical corrections, ReFINE-Lung produces scalable, trustworthy pseudo-labels that reduce false positives and enhance generalizability across clinical domains. This framework lays the foundation for semi-supervised training, domain adaptation, and external benchmarking of lung nodule detection systems in real-world screening environments.

Overview of ReFINE-Lung

Figure. Overview of ReFINE-Lung. ReFINE-Lung is a modular framework designed to curate high-quality pseudo-labels from clinical lung screening CTs. It integrates multi-model candidate generation with ensemble agreement, followed by calibration, unsupervised prediction alignment, prevalence estimation, and cost-aware thresholding. This consensus-based refinement pipeline enhances reliability of predicted nodule probabilities while reducing false positives across diverse clinical domains.

BibTeX

Comming Soon!!!