RPS 1604 - Artificial Intelligence in pulmonary imaging

Author Block: I. Friscia, G. Cicchetti, L. Calandriello, F. Scrocca, L. Cerquiglini, E. De Lorenzis, S. L. Bosello, L. Natale, A. R. Larici; Rome/IT
Purpose: Interstitial lung disease (ILD) is a major complication in systemic sclerosis (SSc) patients. Artificial intelligence (AI) application on high-resolution Computed Tomography (HRCT) has emerged as a tool that may ensure an objective assessment of ILD. The aim of this study is to correlate measures extracted from HRCT images by a deep-learning based software to assess ILD progression and disease-related mortality in SSc patients.
Methods or Background: HRCT scans from a cohort of consecutive SSc-ILD patients at baseline and after 24±3 months were analyzed using AVIEW software (Coreline Soft, South Korea). Quantitative analyses included lung volume, texture, airway, and vascular measurements. Baseline metrics were assessed for their association with ILD progression, defined by criteria based on the INBUILD trial. Changes in AI-derived measurements between two consecutive HRCT scans over the 24-month follow-up were analyzed for their association with SSc-related mortality during the subsequent 36 months.
Results or Findings: A total of 146 HRCT scans from 73 SSc-ILD patients were assessed (mean age 58.4±14.3 years). Thirty-one patients (42.4%) experienced ILD progression over 24 months, which was predicted at baseline by higher percentages of ground glass opacities (GGO) (p=0.05) and reticulation (p=0.05), higher subpleural vessel volumes (p=0.017), and a tendency toward larger distal airways (p=0.066). Serial evaluations demonstrated that progression was associated with a reduction in the percentage of normal lung (p=0.044), and absolute volumes (p=0.009). Patients in the upper quartile for changes in reticular score and airway volume exhibited a higher mortality risk, independently from INBUILD progression (reticular score: OR 3.30, 95%CI 1.03–10.61, p=0.045; airway volume: OR 3.37, 95%CI 1.08–10.51, p=0.036).
Conclusion: Deep learning-based assessment in SSc-ILD identified distinct modifications in lung anatomical components with significant prognostic implications, potentially enabling a precise patient evaluation and stratification.
Limitations: Single center population
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study protocol was approved by the local institutional Committee on Research Ethics

Author Block: S. Kandel, A. Hope, P. Rogalla; Toronto, ON/CA
Purpose: Interstitial lung disease (ILD) is a significant risk factor for radiation pneumonitis in patients receiving pulmonary stereotactic body radiotherapy (SBRT) for cancer. Pre-treatment CT used for planning offer an opportunity to identify patients with ILD prior to therapy. However, manual CT review by thoracic radiologist isn't always available and labor intensive. We developed and clinically implemented a deep learning tool to automatically detect ILD on planning CTs to support risk stratification and workflow efficiency.
Methods or Background: A three-dimensional convolutional neural network (3D VGG16) was trained using 4,393 diagnostic CT scans, including 1,366 ILD cases, normalized for slice thickness and reconstruction algorithm. The algorithm was validated on a cohort of 537 patients treated with SBRT, where ILD prevalence and pneumonitis risk were known. A prospective “silent mode” evaluation was performed on 111 patients prior to live deployment. Two operating thresholds were defined: a low-risk threshold (higher sensitivity) and a high-risk threshold (higher specificity). Clinical integration occurred with automated email notifications to radiation oncologists.
Results or Findings: In the validation cohort, ILD was associated with higher rates of pneumonitis (G2+: 20.5% vs. 5.8%; G3: 10.3% vs. 1%; G5: 2/3 patients had ILD features; all p<0.01). In the silent mode phase, the low-risk threshold flagged 52% of patients (missed 1 ILD case), while the high-risk threshold flagged 14% (missed 1 ILD case). Since clinical implementation, 99 patients have been reviewed, with 16 flagged for ILD. Two previously unrecognized ILD cases were detected through the AI tool, leading to changes in patient management.
Conclusion: An AI-based screening tool for ILD can be integrated into radiation therapy planning workflows, identifying high-risk patients and improving clinical decision-making without delaying treatment. Ongoing work involves expansion to multi-center clinical trials.
Limitations: Single center study.
Funding for this study: None.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: REB approval was obtained.

Author Block: B. K. Budai, A. Wagner, O. Havlicek, P. Konietzke, F. Trudzinski, J. Biederer, C. P. Heußel, H-U. Kauczor, V. Palm; Heidelberg/DE
Purpose: This study aimed to investigate the association between vertebral bone density (T12 BMD) and total thoracic vascular calcification (TTVC) volume, with a special focus on the effect of body composition. Moreover, we aimed to determine whether body composition indices modify the bone-vascular axis in patients with chronic obstructive pulmonary disease (COPD).
Methods or Background: In this prospective multicentric study on 539 COPD patients (COSYCONET Study), chest CT scans were investigated with AI-based tools for T12 BMD, TTVC, and volumetric body composition analysis. Adjusted regression models were constructed to assess the impact of conventional body phenotypes (normal, sarcopenic, non-sarcopenic obesity, and sarcopenic obesity). Stepwise interaction model building included T12 BMD, the intermuscular adipose tissue (IMAT), their interaction, adding BMI, clinical and metabolic covariates, lung function, physical performance, and age.
Results or Findings: A consistent inverse association was observed between T12 BMD and TTVC in all phenotypes, reaching significance in normal nutritional status (β = -0.38, p < 0.01), sarcopenia (β = -0.36, p < 0.01), and non-sarcopenic obesity (β = -0.24, p < 0.01). However, the interaction model for TTVC with T12 BMD could not confirm the conventional phenotype’s significant effect. In a fully adjusted linear regression model, IMAT was identified as an independent predictor of TTVC. Interaction models confirmed age and pack-years as the strongest risk factors of calcification; moreover, IMAT consistently remained a significant independent predictor even in the fully adjusted model (β=0.15, 95% CI 0.015–0.28, p=0.029), while the interaction between T12 BMD and IMAT lost significance only once age was included.
Conclusion: Thoracic IMAT is independently associated with vascular calcification in COPD, although the modifying effect of IMAT on the bone-vascular axis suggests an age-dependent interaction.
Limitations: The absence of a non-COPD control group.
Funding for this study: This study was conducted within the framework of COSYCONET and further partially funded through a collaboration with PERMED-COPD (No: 01EK2203A and 01EK2203B). B.K.B. was supported by the Medical Data Scientist Program of the Medical Faculty of Heidelberg University.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The COSYCONET study was approved by the Ethical Committee of Philipps University Marburg (reference no. AZ 2010-28), as well as the ethics committees of each center.

Author Block: L. Sun¹, G. Bailey¹, B. Rawal¹, C. Marrocchio², P. M. George¹, S. Gerry³, S. R. Desai¹, A. Devaraj¹; ¹London/UK, ²Parma/IT, ³Oxford/UK
Purpose: To evaluate the impact of AI-assisted reading using quantitative CT (qCT) on the performance of thoracic radiologists in identifying clinically significant PPF on serial CT, in patients with marginal forced vital capacity (FVC) decline.
Methods or Background: 102 patients (median age, 60 years [range 34–82]; M=40) with non-IPF fibrotic interstitial lung disease (ILD) with serial CTs >6 months apart and contemporaneous FVC decline of 5.0-9.9% were retrospectively evaluated. Five thoracic radiologists, blinded to clinical data, independently reviewed serial CTs side-by-side, categorising cases as either stable disease (SD) or progressive disease (PD), based on visually estimated changes in ILD extent. Quantitative CT biomarkers of ILD severity were also generated using commercially available AI software (e-Lung, Brainomix). Cases initially categorised as SD, but with a qCT biomarker increase of ≥1.5% on serial CT were re-reviewed in conjunction with software generated fibrosis segmentation overlays, and radiologists then either retained categorisation as SD or changed to PD. Radiologists’ performance with and without AI support was analysed using a Cox proportional hazards model based on progression-free survival (defined as time to death, lung transplantation or 10% FVC decline).
Results or Findings: During a median follow-up of 1085 days (IQR 522–1694), 44 patients died and 3 had lung transplants. Without AI, hazard ratios (HR) for visually-identified PD were 1.83-2.33 (p=0.001–0.017) for all readers. QCT identified 22 to 40 cases per reader for re-evaluation, leading to readers changing PPF categorisation in 45%-94% (n=10-29) of these cases. With AI-assisted reading, HRs increased for all radiologists to 2.34–3.25 (p<0.001–0.002).
Conclusion: In patients with non-IPF fibrotic ILD and marginal FVC decline, AI-based qCT decision support improves reader performance in identifying clinically significant PPF.
Limitations: Retrospective, single-centre study
Funding for this study: Nil sought.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval by Health Research Authority England

Author Block: R. J. PRABAGARAN¹, K. venugopal¹, N. D. Kanase², M. P. Ghate², P. Moorthy¹; ¹Chennai/IN, ²Mumbai/IN
Purpose: To evaluate AI-based ultra-low-dose CT (Delta, 0.5 mm) for lung cancer screening in smokers and non-smokers, validated against histopathology, and compared with a radiologist and machine learning (ML) classifier.
Methods or Background: 179 patients undergoing ultra-low-dose CT chest, all suspicious nodules were biopsied. Histopathology was the reference. AI, radiologist, and ML provided binary malignant/benign predictions. Diagnostic accuracy was assessed per-patient (positive if any malignant lesion) and per-lesion. Sensitivity, specificity, PPV, NPV, and accuracy were calculated with 95% confidence intervals. Subgroup analyses were performed in smokers and non-smokers. False positives were reviewed by histopathology. McNemar’s test compared paired methods.
Results or Findings: Ninety-seven of 179 patients (54.2%) had malignancy. Per-patient accuracy was 81.6% for AI, 88.8% for radiologist, and 87.2% for ML. Sensitivity/specificity were: AI 80.8%/82.9%, radiologist 88.6%/89.0%, ML 87.6%/86.6%. In smokers, radiologist sensitivity/specificity were 86.3%/89.5%; in non-smokers 91.3%/88.6%. False positives were mainly benign fibrosis and granuloma. McNemar’s test showed no significant difference between methods.
Conclusion: AI applied to ultra-low-dose CT demonstrates clinically useful accuracy for lung cancer screening across smokers and non-smokers, but radiologist and ML interpretation remain superior. Novelty lies in combining AI-integrated ultra-thin reconstruction with systematic biopsy validation, highlighting benign mimics as the main limitation. AI holds promise as an adjunct in screening workflows
Limitations: Single-center cohort: Conducted at one institution, which may limit generalizability to broader screening populationsNo longitudinal follow-up: Interval cancers and long-term outcomes were not assessed; thus, the study reflects diagnostic performance at baseline onlyHistopathology spectrum: Benign conditions such as fibrosis, granuloma, and infection were frequent; while biopsy was the reference, some benign diagnoses may evolve over time.
Funding for this study: This research received no external funding and was conducted as part of institutional academic activity.No specific grant from funding agencies in the public, commercial, or not-for-profit sectors was received for this study.”
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Institutional Ethics Committee and written informed consent was obtained from all patients.
This prospective study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board

Author Block: Z. Zhedong; Beijing, China/CN
Purpose: This multicenter study aimed to develop and validate a machine learning model for predicting the growth of pulmonary subsolid nodules (SSN) at different time intervals using CT radiomics. The model is intended to guide personalized follow-up strategies in clinical practice.
Methods or Background: We retrospectively analyzed data from 642 patients with 717 SSNs, collected from three medical centers, who underwent long-term follow-up. Patients were classified into growth and non-growth groups based on SSN growth within 2 or 5 years and were randomly divided into training and internal testing sets. Predictive models were developed using optimal ML algorithms for clinical, radiomics, and clinical-radiomics fusion models to assess SSN growth risk. An independent external test set, including 95 patients with 105 SSNs from a health examination center, was used for validation. Model performance was assessed using the AUC. The SHAP method was used to clarify model rationale.
Results or Findings: XGBoost and LightGBM showed the highest discriminative ability among eight ML models. For 2-year growth prediction, AUCs were 0.823, 0.889, and 0.911 (internal set), and 0.712, 0.734, and 0.734 (external set). For 5-year growth, AUCs were 0.796, 0.838, and 0.849 (internal set), and 0.672, 0.773, and 0.776 (external set). These insights were integrated into a clinical management framework, enhancing clinical utility.
Conclusion: Our interpretable ML model, based on multicenter longitudinal data, accurately predicts SSN changes over 2 years and offers guidance for 5-year follow-up.
Limitations: As a multicenter retrospective study, site-to-site variation in follow-up likely introduced selection and temporal bias and precluded robust subgroup analyses. The exclusively Asian cohort limits generalizability to other ethnicities, underscoring the need for international external validation. Moreover, not all enlarging nodules were pathologically confirmed, indicating that surgical thresholds after progression require further study.
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study received approval from the Institutional Review Boards (IRBs) of three tertiary comprehensive medical centers (No. 2022PHB031-001, No. 2022-0601-01, and No. 2022002).

Author Block: L. Bai¹, S. Xu², A. Li¹, J. Yang¹; ¹Xi'an/CN, ²Shanghai/CN
Purpose: Repeated chest CT is crucial for monitoring children with severe lobar pneumonia, but radiation exposure is a major concern. This study aimed to evaluate the diagnostic feasibility of the ultralow-dose CT at 60kVp with artificial intelligence iterative reconstruction (AIIR) for pediatric pneumonia, compared to routine-dose CT.
Methods or Background: Thirty-three pediatric patients with severe lobar pneumonia (20 boys; mean age, 7.7±3.5years) undergoing follow-up CT within one-week of baseline routine-dose imaging were prospectively enrolled. Routine-dose protocols were age-specific: 100kVp, reference 70mAs (≤5 years); 120kVp, reference 30mAs (6–18 years), and images were reconstructed with hybrid iterative reconstruction (HIR). Follow-up scans used 60kVp, reference 70mAs, and were reconstructed with AIIR. Image noise was defined as the standard deviation of the CT number in chest-wall fat. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were quantified at aortic root. Two blinded radiologists scored anatomical visualization and diagnostic confidence by consensus using 5-point Likert scale (1=poor, 5=excellent). The paired-t test and Wilcoxon signed-rank test were used. The performance of 60kVp-images for demonstrating pneumonia-related imaging findings was evaluated, with routine-dose CT as the reference standard.
Results or Findings: The effective radiation doses were 0.17±0.03mSv at 60kVp and 1.14±0.37mSv at routine dose, respectively (p<0.05). Compared to routine‑dose images, 60kVp-AIIR images demonstrated significantly lower noise, higher SNR and CNR (all p<0.05). Subjective anatomical visualization and diagnostic confidence scores were comparable between 60kVp and routine‑dose scans (4.00±0.74 vs. 4.42±0.67, p=0.13; 4.25±0.62 vs. 4.50±0.67, p=0.51). Sensitivity, specificity, negative predictive value, positive predictive value, and accuracy of 60kVp-scans were 95.45%, 100%, 91.67%, 100%, and 96.30%.
Conclusion: The 60kVp ultralow-dose chest CT with AIIR provides diagnostic image quality comparable to routine-dose CT while reducing radiation exposure by approximately 85%, demonstrating high potential for safely monitoring pediatric lobar pneumonia.
Limitations: Not applicable.
Funding for this study: No funding was received by this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Institutional Review Board approval was obtained.

Author Block: J. Choi, J. Choe, H. J. Hwang, S. M. Lee, E. Chae, J. B. Seo; Seoul/KR
Purpose: In idiopathic pulmonary fibrosis (IPF), coexisting emphysema may attenuate declines in forced vital capacity (FVC), masking progression. Visual thresholds have been used to define combined pulmonary fibrosis and emphysema (CPFE), but no quantitative CT (QCT)–based thresholds have been clinically validated. We aimed to identify a QCT-defined emphysema threshold on HRCT relevant to monitoring disease progression.
Methods or Background: We retrospectively analyzed IPF patients with baseline and 1-year pulmonary function tests. Emphysema and fibrosis extent were quantified using deep learning–based texture analysis. Prognostic value of 1-year decline in FVC (≥5%) and diffusing capacity of carbon monoxide (DLco,≥10%) for survival was assessed across QCT emphysema thresholds using multivariable Cox models. Longitudinal trajectories were modeled with linear mixed-effects analysis. In a subgroup with follow-up HRCT, prognostic association of QCT-fibrosis progression (≥4.52% increase in fibrosis, predefined DLco-anchored threshold) was evaluated in relation to emphysema burden.
Results or Findings: Among 944 patients (mean age, 66.6 ± 7.9 years; 80.6% male), mean QCT emphysema was 1.75±4.29% and fibrosis extent 11.7±9.95%. FVC decline predicted mortality across subgroups based on different emphysema thresholds but showed no significant association in patients with >10% emphysema (HR, 1.54[95%CI: 0.50–4.69]; P=0.45). DLco decline remained robustly associated with mortality regardless of emphysema extent (all, P<0.001). QCT-fibrosis progression also significantly associated with survival across emphysema strata. FVC trajectories differed by the 10% QCT-emphysema threshold (P<0.001), with stability in >10% versus decline in ≤10%. DLco declined in both groups without slope difference (P=0.36).
Conclusion: A QCT-emphysema threshold of ≥10% defines a CPFE-IPF phenotype in which FVC is less sensitive for monitoring progression. DLco decline and QCT-fibrosis extent change may serve as preferred surrogate markers.
Limitations: Retrospective, single-center study; small sample size for HRCT follow-up subgroup analysis.
Funding for this study: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2025-16067456).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This retrospective study was approved by our institutional review board, and the need for written informed consent was waived.

Author Block: D. C. Rotzinger, G. Fahrni, A. Mackowiak, S. Rapacchi, J-B. Ledoux, M. Stuber, C. W. W. Roy, C. Pozzessere; Lausanne/CH
Purpose: To assess the performance of free-running 4D (respiratory motion-resolved) ultrashort-echo-time (UTE) sequences at 1.25mm3 and 0.98mm3 resolution on low-field MRI, compared with commercially available 3DT1VIBE (1.25mm3) and 2DT2HASTE (1.56mm2×6mm) for contrast-free lung imaging.
Methods or Background: We scanned fourteen volunteers on low-field MRI (0.55T). Three experienced radiologists independently rated image quality/artifacts and vessels/airways conspicuity (4-point Likert scale). We calculated median/IQR ratings and assessed differences of pooled and individual ratings via Kruskal-Wallis with Dunn's post-hoc tests. Inter-rater agreement was evaluated with weighted Cohen's Kappa and interpreted according to Landis & Koch (1977). Quantitative analysis included pulmonary artery-to-parenchyma contrast ratio (CR).
Results or Findings: UTE1.25 vs. UTE0.98’s pooled qualitative ratings were not significantly different (p=0.757). However, UTEs ratings (average median±SD: 3.25±0.5) outperformed other sequences; 3DT1VIBE ranked third (3.0±0) and 2DT2HASTE last (2.25±0.5), all p.adj<0.001. When looking at individual rating categories, UTEs excelled especially in vessel (median[IQR]: 4[1]) and airways conspicuity (3[1]), outperforming 2DT2HASTE (2[1.75] and 2[1], p.adj<0.001) and 3DT1VIBE (3[1] and 3[1], respectively, p.adj<0.01). UTEs had milder artifacts than 3DT1VIBE, p.adj<0.05, but were not significantly better than 2DT2HASTE (p.ajd>0.19). Inter-rater agreements were as follows (averages): image quality 0.32 (fair), artifacts 0.15 (slight), vessels 0.50 (moderate), airways 0.64 (substantial). In quantitative analysis, 3DT1VIBE outperformed UTE0.98 (CR 4.4±1.6 vs. 1.92±0.5, p.adj<0.01) and 2DT2HASTE (CR 0.64±0.2, p.adj<0.001), but was not significantly superior to UTE1.25 (CR 2.89±1.1, p.adj=0.09), favoring UTE1.25 over UTE0.98 due to its higher contrast ratio.
Conclusion: UTE1.25 balanced qualitative superiority and quantitative reliability for low-field 0.55T lung MRI, offering superior depiction of vessels and airways compared to 3DT1VIBE and 2DT2HASTE and less artifacts.
Limitations: Pilot study with limited clinical cohort size; no clinical endpoint
Funding for this study: This work was funded by a CHUV Radiology Seed Grant
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: CER-VD, Vaud's Canton indepedant Ethics Committee.