Cancer classification with radiomics in controlled preclinical models

Abstract The premise of radiomics involves extracting high-dimensional quantitative features from medical images to aid clinical decision-making. While radiomics has shown promise in predicting disease characteristics, concerns regarding confounders, reproducibility, and interpretability limit its clinical adoption. In this study, we assessed the ability of radiomic features extracted from contoured CT images to classify two distinct tumour models, CT26 colorectal cancer (CRC) and 4T1 breast cancer (BC), in a highly controlled murine setting. We aimed to provide compelling data for the role of radiomics as a reliable cancer biomarker. We benchmarked radiomics-based classification against previously established blood-based biomarkers, including leukocyte populations and plasma proteins. Feature filtering reduced the original 1409 radiomic features to 18 non-redundant, high-importance predictors, primarily texture-based transformations. Unsupervised clustering via UMAP revealed that radiomics-based features did not segregate tumour types as effectively as blood biomarkers, suggesting potential confounding factors. Supervised machine learning using Random Forest showed that radiomic features achieved a classification accuracy of 0.87, lower than the 0.96 and 0.99 accuracies obtained with cell and plasma biomarkers, respectively. Furthermore, integrating radiomics with blood biomarkers did not enhance classification performance, and feature importance analysis using SHAP identified blood-based markers as the dominant predictors. These findings suggest that while radiomics retains some predictive value, it is less effective than blood biomarkers in this classification task and does not significantly contribute to multimodal tumour classification models. Our study underscores the need for further standardization and validation of radiomics before its clinical implementation.