Exploring free radiomics software tools: a multiparametric evaluation for cancer classification

Abstract

Cancer remains the leading cause of mortality worldwide, accounting for over 10 million deaths annually. Despite recent advances in early detection and screening, traditional diagnostic procedures, such as biopsies, remain invasive and time consuming. Radiomics is a trending method that transforms visual information from medical images into numerical features and offers non-invasive, quantitative analysis of medical images to support clinical decision-making. Numerous radiomics software tools exist in the literature, each offering different features as they include a variety of transformations, calculations, and configurations. Selecting the most suitable tool poses a challenge that requires a comparative analysis due to the lack of standardisation. However, no prior study has systematically compared multiple free radiomics software tools across multiple cancer types while evaluating information fusion strategies, which represents the specific research gap addressed here. This paper aims to evaluate the performance of multiple free and open-access radiomics software tools, analyse the variability in the extracted features and demonstrate that combining different radiomics tools is more appropriate than choosing one individual. This study conducts a comparative analysis of several free and open-access GUI radiomics software tools applied to cancer classification tasks using multiple datasets. The evaluation is carried out considering optimal machine learning pipelines that integrate feature selection and classification components using the radiomic features extracted from the radiomic tools. Information fusion methods are also proposed to demonstrate their high performance against single radiomics tools. Additionally, the radiomic features extracted from each tool were analysed, identifying those most relevant in the classification process. Experiments conducted on 10 datasets and 80 pipeline combinations (10 preprocessing pipelines and 8 classification models) revealed statistically significant differences in F1-score, AUROC and AUPRC a depending on the radiomics tool employed. Importantly, feature fusion consistently outperformed individual tools in classification tasks, reaching improvement of 3% on all the metrics. These results highlight the substantial impact of tool selection on model performance and the enhanced effectiveness of combining features from multiple sources. The study also identified the most relevant features contributing to model performance and the best classification pipeline. This study guides radiologists and researchers in selecting effective radiomics tools and pipelines for cancer classification. The findings underscore the importance of careful radiomics tool selection, as performance varies considerably across software. It demonstrates that fusion-based approaches that integrate features from multiple tools provide more robust and accurate predictions than any single tool.

Ähnliche Arbeiten