Work

1027 Foundation-model-based CT detection of pneumonitis-compatible abnormalities during immune checkpoint inhibitor surveillance

Year: 2026

Type: conference-abstract

Abstract: Background Treatment-attributed pneumonitis is an important toxicity of immune checkpoint inhibitor (ICI) therapy, requiring vigilant on-treatment surveillance. While baseline CT-based risk models exist, 1 2 tools for on-treatment surveillance remain limited. We developed an automated CT image model to detect pneumonitis-compatible abnormalities.Methods We retrospectively assembled 11,611 thoracic CT scans paired with radiology reports from 1,558 patients with advanced solid tumors (NSCLC, melanoma, RCC, CRC) across 7 US and EU institutions. The primary endpoint, pneumonitis-compatible abnormality, was labeled by LLM-assisted review of the radiology report and defined as any active parenchymal pattern consistent with pneumonitis (e.g., ground-glass opacity, consolidation, organizing-pneumonia-like), irrespective of cause. The secondary endpoint, treatment-attributed pneumonitis (ICI-, radiation-attributed, or combined), additionally incorporated longitudinal context and medication, and indeterminate cases were excluded. A pre-specified ICI-treated NSCLC test set (299 patients) was held out from model development and cross-validation (1,259 patients; CV set).Per-scan features from three frozen pretrained encoders (FMCIB,3 CT-FM,4 and CT Response Score (CTRS))5 were concatenated into a 4,736-dimension vector. An image-text contrastive projection (InfoNCE) was trained on the CV set to align image features with paired report embeddings from a frozen biomedical text encoder.6 Logistic-regression probes and thresholds targeting sensitivity ≥0.80 were tuned using 5-fold patient-stratified cross-validation on the CV set. Test data were excluded from all training and tuning, and test inference was fully automated and used image embeddings only.Results Pneumonitis-compatible abnormality and treatment-attributed pneumonitis were evaluable in 4,385 and 4,194 development CT studies (17.2% and 2.6% positive) and 1,065 and 983 test CT studies (24.2% and 3.8% positive). After study-level aggregation, pneumonitis-compatible abnormality AUROC was 0.68 (95% CI, 0.66-0.70) in the CV set and 0.69 (95% CI, 0.65-0.72) in the test set. For treatment-attributed pneumonitis, AUROC was 0.73 (95% CI, 0.69-0.78) and 0.74 (95% CI, 0.66-0.83) in the CV and test set. At the pre-specified threshold for pneumonitis-compatible abnormality, test sensitivity was 0.86 and NPV 0.88, with specificity 0.34 and PPV 0.29.Conclusions A scalable, fully automated CT image-only model generalized to a held-out ICI cohort for both a broad pneumonitis-compatible abnormality and a stricter treatment-attributed endpoint. At the pre-specified threshold, the model showed high sensitivity and high NPV, supporting further validation as an adjunct for longitudinal ICI safety monitoring, where a negative read may support continued treatment and a positive read may prompt clinical review. Future work will include refining operating thresholds and obtaining adjudicated labels to evaluate ICI- vs. radiation-attributed pneumonitis subsets.References Muneer, et al. CIPHER foundation model for pre-ICI CT prediction of immune checkpoint inhibitor pneumonitis. medRxiv. 2026. doi:10.64898/2026.04.21.26351428.Smith, et al. CT radiomics and deep learning prediction of immune checkpoint inhibitor pneumonitis. JCO Clin Cancer Inform 9. 2025;e2400198.Pai S, Bontempi D, Hadzic I, et al. Foundation model for cancer imaging biomarkers. Nat Mach Intell. 2024;6:354-367.Pai S, Pillai SP, Bontempi D, et al. CT-FM: a foundation model for computed tomography. arXiv:2501.09001. 2025.Sako C, Duan C, Maresca K, Kent S, et al. Real-world and clinical trial validation of a deep learning radiomic biomarker for PD-(L)1 immune checkpoint inhibitor response in advanced non-small cell lung cancer. JCO Clin Cancer Inform. 2024;8:e2400133.Boecking B, Usuyama N, Bannur S, et al. Making the most of text semantics to improve biomedical vision-language processing. European Conference on Computer Vision (ECCV). 2022. Lecture Notes in Computer Science, vol 13696. Springer, Cham.Ethics Approval Ethics approval for US data: The study was conducted under IRB-approved procedures using de-identified data for patients diagnosed with advanced cancer and treated between Jan. 1, 2017 and December 31, 2023. All records were de-identified per HIPAA guidelines at the institution level. Upon transfer, the data was quarantined and then re-inspected by authorized personnel prior to ingestion to ensure compliance and that no PHI was present in the records. Ethics approval for EU data: The study was conducted under IRB-approved procedures using de-identified data for patients diagnosed with Stage-IV NSCLC and treated between Jan. 1, 2017 and December 31, 2021. All records were de-identified per GDPR requirements at the institution level. The patients were also notified that their de-identified data would be part of a study and were given the required time and opportunity to respond if they had any objection. Upon transfer, the data was quarantined and then re-inspected by authorized personnel prior to processing to ensure compliance and that no PHI was present in the records.

Source: Journal for ImmunoTherapy of Cancer

DOI: 10.1136/jitc-2026-sitc2026.1027