Radiomics-based outcome prediction for irinotecan-TACE in colorectal liver metastases: advanced analysis from the prospective CIREL trial

Fecha de publicación: Fecha Ahead of Print:

Autores de IIS La Fe

Participantes ajenos a IIS La Fe

  • Bodalal Z.
  • Mendoza Ferradás F.J.
  • Maxouri O.
  • Iezzi R.
  • Gjoreski A.
  • Spiliopoulos S.
  • Bansaghi Z.
  • Gonçalves B.
  • Zeka B.
  • Kaufmann N.
  • Taieb J.
  • Beets-Tan R.
  • Pereira P.L.

Grupos

Abstract

Objectives: Transarterial chemoembolization (TACE) is a promising locoregional therapy for unresectable colorectal liver metastases, but patient selection remains challenging. We aimed to develop and validate prognostic radiomics-based machine learning models in a multicenter, prospectively collected drug-eluting microsphere TACE cohort. Materials and methods: We retrospectively analyzed 76 patients (176 lesions) from the prospective CIREL registry trial. Radiomic features were extracted from each lesion. We tested three types of imaging markers: general radiomics, intensity-based features, and lesion volume. For each, we derived baseline and delta features, reflecting the difference in feature vector values between baseline and first follow-up. Using a center-based split, we trained genetic/evolutionary machine learning models to predict survival and lesion-level response. Results: The median age of the final study population with baseline imaging was 66 years (IQR, 59–71), with 67.1% (n = 51) of patients identifying as male. On external validation, the baseline intensity algorithm was the only significant survival-prediction model (AUC = 0.79, 95% CI = 0.57–0.95; p = 0.011), outperforming baseline radiomics (AUC = 0.69, 95% CI = 0.47–0.86; p = 0.100) and baseline volume (AUC = 0.56, 95% CI = 0.37–0.74; p = 0.574). Radiomic prediction models stratified patients into distinct overall survival risk groups, with low-risk patients showing a median survival of 696 days versus 453 days (log-rank p = 0.0267). Integrating imaging features with laboratory variables improved lesion-level response assessment (AUC = 0.86, 95% CI = 0.66–0.99; p = 0.006), but did not enhance OS prediction. Lesion-level response was best identified by delta radiomics (AUC = 0.83, 95% CI = 0.63–0.97; p = 0.008). Conclusion: Radiomics-based machine learning models could predict overall survival in patients treated with irinotecan-TACE, offering a potential tool for patient selection. Key Points: Question Can radiomics and machine learning predict outcomes in patients with colorectal liver metastases treated with irinotecan-TACE, aiding in patient stratification and selection? Findings Baseline intensity features predicted overall survival (AUC = 0.79), while delta radiomics identified lesion response (AUC = 0.83) in a multicenter cohort. Clinical relevance These models can help identify patients likely to benefit from irinotecan-TACE and lesions most responsive to treatment. Further development would enable personalized therapy that may improve survival and reduce unnecessary interventions in non-responders. © The Author(s) 2026.

Datos de la publicación

ISSN/ISSNe:
0938-7994, 1432-1084

EUROPEAN RADIOLOGY  SPRINGER

Tipo:
Article
Páginas:
6981-6996
PubMed:
42053592
Enlace a otro recurso:
www.scopus.com
Factor de Impacto:
1,730 SCImago ℠
Cuartil:
Q1 SCImago ℠

Citas Recibidas en Web of Science: 1

Documentos

  • No hay documentos

Métricas

Filiaciones mostrar / ocultar

Keywords

  • Colorectal liver metastases; Machine learning; Radiomics; Survival prediction; Transarterial chemoembolization

Cita

Compartir