Back to 2025 Abstracts

---

MRI-BASED CLINICAL-RADIOMICS ANALYSES OF COLORECTAL LIVER METASTASES TO PREDICT TUMOR GROWTH PATTERNS BY MACHINE LEARNING APPROACH.
Simone Famularo^*2,3, Elena H. Tran², Angela Ammirabile¹, Luca Boldrini², Cristiana Bonifacio¹, Caterina Mele², Edda Boccia², Enza Genco², Guido Costa¹, Felice Giuliante², Guido Torzilli¹, Luca Viganò⁴, Francesco Ardito^2
1Dept. of Surgery, Hepatobiliary and general surgery division, IRCCS Humanitas Research Hospital, Rozzano, Lombardia, Italy; ²Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Roma, Lazio, Italy; ³Institut de Recherche Contre les Cancers de l'Appareil Digestif, Strasbourg, Grand Est, France; ⁴Cliniche Gavazzeni SpA, Bergamo, Lombardia, Italy

Introduction
Tumor growth patterns in colorectal liver metastases (CRLM) were associated with different recurrence risk after liver resection, but this information is not available before therapy decision. This study explores a combined clinical-radiomic approach to perform MRI analysis of hepatic metastases with the aim to predict tumor growth patterns by different machine learning algorithms, specifically distinguishing between replacement and desmoplastic/pushing types.
Methods
3-phases MRI scans were retrospectively collected among 2 Italian centers. DICOM files were manually segmented to detect the tumor (core) and the peritumoral area (ring). Clinical variables association with the outcome (replacement vs. desmoplastic/pushing pattern) was assessed via statistical analysis. Radiomics features were extracted in the hepatobiliary phase using Pyradiomics and ComBat harmonization was applied to features to reduce inter-centers variability effects. Features selection was performed via minimum redundancy maximum relevance (MRMR) approach. Models of different complexity, including logistic regression, Random Forest, XGBoost and SVM were evaluated. The area under the curve (AUC) of the receiver operating characteristics (ROC), together with sensitivity and specificity were computed to assess model performance.
Results
Two hundred-twelve patients were enrolled, dividing them into training (148 patients, 70%) and test (64 patients, 30%) cohorts. Harmonization via ComBat reduced feature instability related to imaging protocol variances, enhancing the overall reliability of the feature set. Feature selection on the training set yielded 1 clinical variable (tumor size) and 5 radiomics features. The Random Forest model showed the best performance with the combined clinical-radiomic approach, yielding an AUC of 0.91, sensitivity of 0.87, and specificity of 0.79 on the training set. On the test set, these metrics were lower (AUC 0.70, sensitivity 0.73, specificity 0.60, figure 1), demonstrating a good calibration (figure 2).
Conclusions
The clinical-radiomic modeling approach shows promising potential for accurately predicting growth patterns of hepatic metastases in MRI. Future studies will compare the combined clinical-radiomic approach with the purely radiomic and clinical ones.

---

Back to 2025 Abstracts