Identifying Genetic Mutation Status in Patients with Colorectal Cancer Liver Metastases Using Radiomics-Based...

Area of Science:

• Oncology
• Radiology
• Medical Imaging
• Machine Learning

Background:

• Genetic mutation status, particularly KRAS, is crucial for treatment selection and prognosis in colorectal cancer liver metastases (CRLM).
• Radiomics, the extraction of quantitative imaging features, offers a non-invasive approach to assess tumor characteristics.

Purpose of the Study:

• To investigate the association between radiomics features from CT scans and KRAS mutation status in CRLM patients.
• To validate machine-learning models for predicting KRAS mutation status using radiomics in an independent external dataset.

Main Methods:

• Radiomics features were extracted from semi-automatically segmented CRLM in pre-treatment CT scans from the CAIRO5 trial (n=255).
• Machine learning classifiers (Random Forest, Gradient Boosting, LightGBM, Ensemble) were trained to predict KRAS mutation status.
• Model performance was evaluated internally and externally using the Netherlands Cancer Institute (NKI) dataset (n=129).

Main Results:

• Internal validation showed good performance with AUC scores ranging from 0.72 to 0.86 for the machine-learning models.
• External validation demonstrated significantly poorer performance, with AUC scores between 0.47 and 0.56.
• The study highlights a discrepancy between internal and external validation results.

Conclusions:

• Machine-learning models using CT radiomics can identify KRAS mutation status in CRLM with good internal accuracy.
• Poor performance in external validation underscores the critical need for robust external validation of radiomics models.
• Mandatory external validation is essential for assessing the clinical applicability of radiomics in future CRLM research.