# Automated interstitial lung abnormalities detection at CT: external validation and potential recognition of traction bronchiectasis/bronchiolectasis

## TL;DR

This study validates an AI system for detecting lung abnormalities in CT scans and shows its scores correlate with the severity of bronchiectasis.

## Contribution

The study externally validates an AI system for interstitial lung abnormalities and links its outputs to bronchiectasis severity.

## Key findings

- The AI system achieved high accuracy (AUC 0.841 and 0.823) in detecting ILA across two different populations.
- AI scores correlated strongly with the severity of traction bronchiectasis/bronchiolectasis, especially in severe cases.
- The system's performance was consistent across diverse populations, suggesting robust generalizability.

## Abstract

An artificial intelligence (AI) system for detecting interstitial lung abnormalities (ILA) was previously developed but requires external validation. This study aimed to examine the robustness across different populations and investigate associations between the system outputs and traction bronchiectasis/bronchiolectasis severity patterns.

CT scans from population-based samples of the Rotterdam Study (2018–2019) and the Age Gene/Environment Susceptibility Reykjavik (AGES-Reykjavik) Study (baseline CT: 2002–2006, follow-up CT: 2007–2011) were used in this secondary analysis of the two cohorts. The AI system calculated ILA probability score (AI score) in the range from 0 to 1. Three experienced readers evaluated independently all CT scans for ILA, and two chest radiologists assessed traction bronchiectasis/bronchiolectasis using the 4-scale traction bronchiectasis/bronchiolectasis index (TBI) for severity by consensus. Receiver operating characteristic (ROC) analysis and Kruskal–Wallis test were used for statistical analysis.

The system analyzed 932 CT scans of the Rotterdam Study (mean participant age, 79.6 years ± 4.3 (SD), 482 women) and 5242 CT scans of the AGES-Reykjavik Study (mean participant age, 76.4 years ± 5.6, 3032 women), and achieved area under the ROC curve of 0.841 (95% CI 0.804, 0.879) and 0.823 (95% CI 0.798, 0.847), respectively. AI scores correlated with readers’ certainty, decreasing from unanimous ILA cases to No-ILA cases. Higher baseline AI scores correlated with greater severity of traction bronchiectasis/bronchiolectasis (TBI-3: 0.931 [IQR, 0.911–0.932], TBI-2: 0.738 [IQR, 0.406–0.880], TBI-1: 0.537 [IQR, 0.317–0.761], TBI-0: 0.250 [IQR, 0.136–0.455]).

The system demonstrated robust ILA detection performance across different populations, with AI scores showing associations with traction bronchiectasis/bronchiolectasis severity.

The online version contains supplementary material available at 10.1007/s11604-025-01917-z.