Scalable Detection of Undiagnosed ILD in Population Screening: A Multi-Cohort Study using 3D Foundation Models

Niccolò McConnell, Mehran Azimbagirad, Daryl O. Cheng, Daisuke Yamada, Ryoko Egashira, Robert Chapman, John McCabe, Shanshan Wang, David Lynch, Greg Kinney, Pardeep Vasudev, SUMMIT Consortium, Paul Taylor, Daniel C. Alexander, Sam M. Janes, Joseph Jacob

Keywords: Foundation Model, Large-scale CT data, UILD, ViT, ConvNeXt

TL;DR: Large-scale validation of 3D foundation-model CT transformers for detecting subtle undiagnosed ILD in CT screening cohorts.

Abstract: Undiagnosed interstitial lung disease (UILD), an early form of lung fibrosis is increasingly detected in population-based low-dose computed tomography (LDCT) screening but remains systematically under-reported due to its subtle appearances. We develop and validate a foundation-model–augmented deep learning system for UILD detection across two of the largest thoracic CT cohorts worldwide: SUMMIT, the UK’s largest LDCT screening study (>11,000 scans), and COPDGene, a multi-centre US cohort spanning 21 scanners and >8,800 scans. We propose ViT-3D-TE, a multi-token 3D Vision Transformer that aims to preserve both high-frequency focal texture and diffuse parenchymal change through CLS, MAX, and AVG token fusion. The model is initialised with TANGERINE, an open-source 3D masked autoencoder pretrained on 98,000 full-volume LDCT scans, providing volumetric priors essential for stable optimisation. Trained solely on SUMMIT and evaluated on COPDGene without domain adaptation, ViT-3D-TE achieved strong performance (AUROC 0.9805, AUPRC 0.7699 internal; AUROC 0.9705, AUPRC 0.6170 external), representing 17x and 25x improvements over random baselines at clinically realistic cohort prevalences (4.6% and 2.5%). We further introduce ConvNeXt-2.5-MIL, a slice-based 2.5D alternative that performs competitively without relying on 3D foundation model pretraining. Together, these results provide, to our knowledge, the largest real-world validation to date of deep learning for UILD detection and demonstrate that foundation-model–enhanced 3D Transformers offer a practical and scalable pathway for integrating UILD detection into national LDCT screening workflows.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Foundation Models

Registration Requirement: Yes

Reproducibility: https://github.com/niccolo246/UILD-detection-deep-learning

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 36