RPS 304 - Hot Topic: AI-driven lung cancer screening

Author Block: D. C. Rotzinger, C. Pozzessere, G. Fahrni, T. Saliba, C. Chevallier, Y. Siddiki, V. Vitzthum, C. Von Garnier, L. Gallego Manzano; Lausanne/CH
Purpose: To qualitatively and quantitatively evaluate the European Society of Thoracic Imaging (ESTI) lung cancer screening (LCS) low-dose CT protocol’s image quality of the at three weight-based CTDIvol levels, and to assess the effect of participant size on image quality.
Methods or Background: ASiR-V80 lung-kernel reconstructions of 173 LCS participants (69 females, mean age 62.9 years) were retrospectively analyzed, and size metrics (weight, thoracic perimeter) obtained. Three readers independently rated diagnostic quality, noise and sharpness on a 5-point Likert scale. Mean, median, standard deviation, categorical percentage, and inter-reader agreement were calculated. Dose-normalized contrast-to-noise ratio (CNR/CTDIvol) was evaluated as a function of size by linear regression.
Results or Findings: Mean weight and thoracic perimeter were 78kg (43-180) and 107cm (81-142.6). Median CTDIvol was 0.40mGy (0.40–0.51, n=5), 0.85mGy (0.60–2.95, n=96), and 1.69mGy (0.86–3.06, n=72). Reader scores (mean/median/SD) were 4.02/4/0.73 for quality, 3.95/4/0.73 for noise, and 4.06/4/0.72 for sharpness. Quality was rated good/excellent in 77.8% of cases, moderate in 20.4%, and poor/non-diagnostic in 1.7%. Noise and sharpness were rated minimal/excellent in 67.6% and 78% of cases, respectively. Inter-reader agreement was 59.0% for quality, 47.4% for noise, and 58.4% for sharpness. CNR/CTDIvol decreased with increasing weight and perimeter (p≤0.001), with perimeter explaining more variance in heavier patients (R² 0.432 vs 0.168 in >80kg; 0.250 vs 0.291 in 50–80kg). At the conservative 25th‑percentile, log-CTDI models suggested dose tiers of ~0.4mGy at <63kg/<30cm, 0.9mGy at 87kg/34cm, and 1.6mGy at 98kg/36.3cm.
Conclusion: The ESTI LCS protocol provides high-quality images with potential for further dose reduction. Thoracic perimeter appears more robust than weight in balancing image quality and dose, particularly in heavier patients.
Limitations: Retrospective, single‑center design; small <50kg subgroup sample size; single reconstruction algorithm
Funding for this study: No external funding
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics Committee of the Canton de Vaud - CER-VD

Author Block: R. Mura¹, A. Kerpel-Fronius², T. G. Blum³, J. Chorostowska-Wynimko⁴, P. Zolda¹, H. Prosch¹; ¹Vienna/AT, ²Budapest/HU, ³Berlin/DE, ⁴Warsaw/PL
Purpose: To explore the current status of lung cancer screening (LCS) training practices, identify existing gaps, and further define key competencies to be included in LCS educational curricula.
Methods or Background: As part of the European SOLACE project, a structured cross-sectional survey, consisting of 11 closed- and open-ended items, was designed and administered to a panel of European LCS experts. Participants were contacted via email and invited to individual Zoom interviews (May-August 2025), each lasting approximately 30 minutes. Descriptive statistics were used for data analysis.
Results or Findings: Seventeen LCS experts were interviewed (response rate 85%, 17/20) from 11 European countries, including 10 radiologists, 4 pulmonologists, 2 thoracic surgeons, and 1 project manager. Reported training activities ranged from established programs (4/11 countries, mandatory in 2/4), to learning initiatives (4/11) and/or planned programs (4/11). Radiologists (100%), pulmonologists (71%), and thoracic surgeons (67%) were identified as the main target groups for training. On a 6-point Likert scale (0= not important, 5= extremely important), experts indicated limited awareness of training needs as the most relevant gap (mean score 3.1±2.2). Key educational areas to be strengthened included management of incidental findings (4.2±0.6), collaborative (4.1±1.2) and communicative skills (3.8±1.5). Among essential competencies for professionals involved in LCS, the top-rated were familiarity with guideline-based nodule management (4.7±0.5), proficiency in managing of incidental findings (4.5±0.6), and basic knowledge of AI-tools (4.5±0.8).
Conclusion: The findings reveal a highly heterogeneous LCS training landscape across Europe, further challenged by limited awareness of training needs, and highlight the importance of developing shared European curricula. Priorities include technical competencies as well as collaborative and communication skills.
Limitations: Not applicable.
Funding for this study: The SOLACE project is co-funded under the EU4Health Programme 2021–2027 under grant agreement no. 101101187. Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or HaDEA. Neither the European Union nor the granting authority can be held responsible for them.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: K. Kvaková¹, D. Kvak², J. Dandár², M. Stastny², J. Olejko³, J. Kultan³, J. Gerold⁴, A. Kotoucova⁵, C. Vojtek⁶, P. Struna⁷, E. Dvorackova⁸, J. F. Smetana⁸; ¹Brno/CZ, ²Prague/CZ, ³Olomouc/CZ, ⁴Sumperk/CZ, ⁵Benešov/CZ, ⁶Nový Jičín/CZ, ⁷Slaný/CZ, ⁸Kladno/CZ
Purpose: Although AI systems for chest X‑ray (CXR) interpretation may enable earlier detection of thoracic malignancy, large‑scale real‑world evidence remains limited. We present results of a joint Carebot–Bristol Myers Squibb program integrating AI decision‑support into routine workflows to evaluate feasibility, flagged‑case yield, and downstream actions for suspected malignancy (primary lung cancer or pulmonary metastases).
Methods or Background: All CXRs across nine Czech hospitals (regional and tertiary; two pneumo‑oncology centers) from Jan 1–Jun 30, 2025 were automatically analyzed with commercial‑stage AI software (Carebot AI CXR; Carebot s.r.o., Czechia). The AI flagged abnormalities; a joint panel of hospital and interim Carebot radiologists classified exams as suspicious for thoracic malignancy. Suspicious cases entered fast‑track diagnostics. Primary outcomes: (i) proportion flagged as suspicious, (ii) patients referred for diagnostic work‑up, and (iii) confirmed malignancies.
Results or Findings: Among 96 459 CXRs, the algorithm identified abnormalities in 16 030 exams (16.6%). Multidisciplinary review classified 837 CXRs (0.87%) as suspicious. Follow‑up status was available in 561/837 (67.0%): 211 (25.2% of all suspicious; 37.6% of those with status) were referred for immediate diagnostic work‑up, 350 (41.8% of all; 62.4% of recorded) were cleared without further work‑up, and 276 (33.0%) remain under evaluation. Among the 211 investigated, 54 previously undiagnosed thoracic cancers were detected, 70 confirmed known malignancies or pulmonary metastases, 38 await confirmation of nodule origin (biopsy or PET/CT), 20 entered long‑term radiologic surveillance, and 29 were recalled owing to missing clinical data.
Conclusion: In this single‑country, real‑world, multicenter analysis of routine CXR exams during H1‑2025, the Carebot–BMS AI triage workflow enabled early identification of thoracic malignancies within routine care while triaging <1% of exams for expedited review. Ongoing follow‑up will further characterize the scalability of this AI‑supported workflow.
Limitations: This observational study lacks a contemporaneous control.
Funding for this study: Carebot s.r.o. and Bristol Myers Squibb supported software integration and coordination; no public grant funding was received.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: N. Shabshin¹, Y. Kimmel², M. Armoni³, Y. Schiffenbauer², A. Grubstein³, L. Roshkovan¹, A. Iannessi⁴, A. F. Nitu⁵, E. Atar³; ¹Philadelphia, Afula/IL, ²Petah Tikva/IL, ³Petach Tikva/IL, ⁴Nice/FR, ⁵Bucharest/RO
Purpose: Lung Cancer Screening (LCS) programs are expanding worldwide. However, barriers including high costs, radiologist shortages, and radiation concerns remain. Within LCS populations, 84% present Lung-RADS 1,2, requiring 12-month follow-up, while scores 0,3,4 require closer monitoring. Chest Digital Tomosynthesis (DTS) offers lower radiation, faster reading, and reduced costs compared to CT. Prior studies found DTS may have a potential role in LCS. This study aims to evaluate whether cold-cathode DTS is comparable to CT when classifying patients to 2 groups: A:12-month follow-up (Lung-RADS 1,2) vs. B closer follow-up/workup (Lung-RADS 0,3,4).
Methods or Background: A cohort of 38 patients with a diagnostic chest CT underwent a supine DTS (Nanox.ARC). Three radiologists reviewed the DTS and CTs blindly and independently maintaining a one-month gap between modalities. Patients were classified into groups A and B. The classification was compared between DTS and CT.
Analyses included negative and positive predictive value (NPV, PPV), inter/intra-reader agreement, and concordance between DTS and CT.
Results or Findings: On both CT and DTS, 25 patients were categorized as group B and 13 as group A. None of the patients migrated from one group to the other when compared to CT and therefore DTS and CT showed no difference in follow-up recommendations. NPVs ranged 72.7–90.0% and PPVs 87.5–92.0%. Inter-reader agreement for DTS was comparable to CT (91.4%, 82.9%, 88.6% vs. 88.6%, 82.9%, 88.6%). Inter- and intra-reader agreement were strong, with kappa values >0.65.
Conclusion: Preliminary results suggest DTS can classify patients to those who can continue with 12-months follow-up (Lung-RADS 1–2) and others that require closer follow-up/workup. DTS can potentially reduce the use of CT and thus reduce radiation, costs, and reading time in LCS..
Limitations: small population of clinical patients, not LCS patients
Funding for this study: Study was funded by Nano-x Imaging
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB approvals were granted

Author Block: L. Gozzi, I. Cecchi, D. Cozzi, E. Cavigli, G. Picozzi, G. Gorini, A. Bindi, C. Moroni, V. Miele; Firenze/IT
Purpose: To investigate the association between emphysema, assessed both visually and quantitatively, socio-demographic, smoking, clinical variables, and radiological signs of lung cancer (LC) and coronary artery calcifications (CAC), in a population of 663 patients enrolled in the Tuscan Italung2-CCM lung cancer screening (LCS) programme.
Methods or Background: We included 663 patients aged 55–75 years, with ≥30 pack-years smoking history. Baseline low-dose CT (LDCT) was evaluated both visually for emphysema/CAC and quantitatively using Pulmo 3D Syngo.Via software (LV, MLD, LAV, P15).
Results or Findings: Among the participants, 96.1% had negative/downgraded LDCT findings, while 26 people (4%) were classified as Lung-RADS 4B/4X; 16 of them were confirmed lung cancers. Emphysema was visually detected in 43% of participants and was significantly more frequent in men (OR=1.60), heavy smokers ≥ 40 pack-years (OR=1.96), and LC patients (OR=8.76). Quantitative emphysema values correlated with male sex, age and former smoker status. CAC was present in 67% of scans, more common in men, older participants, and smokers, but showed no independent association with LC and emphysema. Visual scoring detected more emphysema cases (288) compared to software-based quantification (171), likely due to the masking effect of smoking-induced bronchiolitis on density-based measurements.
Conclusion: This study confirms emphysema as an independent predictor of LC, while CAC reflects shared cardiovascular risks without independent LC association. Visual evaluation outperforms quantitative assessment in detecting emphysema, especially in current smokers, highlighting the need to integrate standardised evaluation of emphysema, CAC, and smoking history in LCS lung programmes.
Limitations: The limitations of the study are inter-scanner variability across centres, potentially affecting CT measures. Moreover, focusing our analysis purely on subjects enrolled in the ITALUNG2-CCM cohort limits the generalisability of its findings. Future studies should harmonize protocols and expand recruitment for broader external validity.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: J. Jahn, R. T. Schirrmeister, F. Bamberg, J. Weiß; Freiburg im Breisgau/DE
Purpose: Altered body composition (BC), such as reduced muscle mass or excess adipose tissue, is an independent predictor of cardiovascular and cancer-related mortality. It remains unclear whether deep learning can estimate mortality risk directly from chest CT slices. Here, we developed a deep learning model (CTchest-risk) to estimate mortality risk from lung screening chest CTs in heavy smokers.
Methods or Background: Data from the CT arm of the National Lung Screening Trial (NLST) were used. CTchest-risk was developed in two steps: (1) segmentation of skeletal muscle (SM), subcutaneous adipose tissue (SAT), and intramuscular adipose tissue (IMAT); (2) training of a second model using segmentation masks combined with five axial slices centered at the thoracic vertebra T5 to predict mortality. The training dataset included 14,208 participants (39,304 individual scans). Independent testing was performed on 8,971 individuals not used for development. Primary endpoint was 6-year mortality. CTchest-risk was categorized into three groups (<15%; 15–85% [reference]; >85%). Survival was assessed with Kaplan–Meier analysis and Cox regression adjusted for age, sex, race, BMI, smoking status, diabetes, hypertension, and history of stroke or heart disease. Model performance was evaluated with Harrell’s C-index and nested likelihood-ratio tests (LRT).
Results or Findings: In the test cohort (mean age 61.3±5 years; 41.0% female), 462 deaths (5.1%) occurred over a median follow-up of 6.5 years. Kaplan–Meier estimates showed a graded association between CTchest-risk groups (log-rank p<0.0001). High-risk individuals (>85%) had 38% higher adjusted mortality compared with the reference group (aHR 1.38, 95% CI 1.11–1.71, p=0.004). In nested model comparison, adding CTchest-risk to clinical factors improved prognostic performance (C-index 0.725 vs. 0.731; LRT p<0.001).
Conclusion: CTchest-risk identifies high-risk individuals from lung screening chest CT and predicts mortality beyond traditional risk factors in heavy smokers.
Limitations: No external validation set.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: All NLST-participants provided written informed consent for the original trial and secondary use of the data was approved by the local IRB.

Author Block: H. Huang; China/CN
Purpose: The differentiation between focal pneumonia-type lung cancer (F-PTLC) and focal pulmonary inflammatory lesions (F-PIL) presents significant diagnostic challenges due to their overlapping imaging features, often leading to delayed or misdirected clinical interventions. This research develops an artificial intelligence (AI) model that integrates clinic-radiological data, radiomics features, and deep learning features to distinguish between F-PTLC and F-PIL.
Methods or Background: We retrospectively analyzed 299 patients across two centers, splitting them into training (n = 213) and validation (n = 86) datasets. Final radiomics features and deep learning (DL) features were extracted and optimised separately using PyRadiomics and four pre-trained convolutional neural networks (CNNs), thereby providing the foundations for the radiomics (Rad) model and the DL model. Researchers have developed an AI-based diagnostic model by combining radiomic signatures, deep learning algorithms and clinical-radiological data. The model's effectiveness was rigorously assessed through a comprehensive five-fold cross-validation process.
Results or Findings: Univariate and multivariate logistic regression analyses revealed that gender, CEA levels, and CT mean are significant predictive factors. The performance of this AI model surpasses that of all individual models in both the training and testing datasets, with AUC values of 0.917 and 0.845, respectively. Grad-CAM pinpointed critical image areas impacting judgments, whereas SHAP assigned precise weight values to individual predictors, thereby boosting model clarity. The results of the Hosmer-Lemeshow (HL) test indicate a satisfactory fit of the model, while Decision Curve Analysis (DCA) further confirms the significant advantage of the combined model in clinical practice.
Conclusion: This study's multimodal AI system dramatically improved diagnostic accuracy in differentiating F-PTLC from F-PIL cases by combining clinic-radiological data, radiomic analysis, and deep learning algorithms. The integrated approach offers clinicians a robust framework for delivering personalized, precision-based medical care.
Limitations: Not applicable
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: R. C. Yarol¹, S. Gezer¹, E. Unel², D. Unay¹, O. Dicle¹; ¹Izmir/TR, ²İzmir/TR
Purpose: This study aimed to develop a deep learning based decision support system to
assist radiologists in the detection, follow-up, and diagnostic evaluation of nodules on
thorax CT images.
Methods or Background: Patients who underwent thorax CT and had at least one pulmonary nodule
between January 2022 and December 2023 at one institution were
retrospectively screened. A dataset consisting of 2,152 nodules from 205 patients was
compiled. In addition to nodule annotations, the location (intrapulmonary, subpleural,
perivascular) and density (solid, semisolid, ground-glass) of each nodule were
recorded for analysis. The dataset was divided into training and testing subsets.
A 2D U-Net-based model was employed for lung and lobar segmentation.
Vessel structures were labeled using a thresholding method. A 3D U-Net model was
utilized for nodule detection and segmentation. The model’s performance on the test
set was evaluated using sensitivity, specificity, and F1-score. Causes of false positives
and false negatives were retrospectively analyzed.
Results or Findings: The developed model achieved a sensitivity of 0.979, specificity
of 0.774, and F1-score of 0.864. The models sensitivity for nodules larger than 6 mm was 92.6%, precision was 74.3%, and F1-
score was 0.824. For nodules smaller than 6 mm, the sensitivity was 98.8%, precision
was 77.9%, and F1-score was 0.871.
The most common cause of false positives (48.8%) was found to be subpleural
fibrotic strands. In the analysis of false negatives, there was a statistically significant
association between nodule density and false-negative predictions by the model
(p<0.001). No statistically significant relationship was found between nodule location
and false-negative results.
Conclusion: Our deep learning model demonstrated high accuracy in pulmonary nodule detection, highlighting its potential to support radiologists, though further multicenter validation is warranted.
Limitations: Retrospective nature
Funding for this study: This study is funded by TUBITAK and X- Focus AI
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved by Ethics Committee (6779-GOA)