Roadmap: From Clinical Need to Translatable Tool

摘要

Learning Objectives: To define biomarkers. To determine the steps in a radiomics study. To describe analytical and clinical validation Body: This machine learning talk will be delivered through the lens of clinical radiology. The most relevant imaging biomarkers are diagnostic, prognostic, monitoring and predictive. Radiomics methodology allows imaging biomarkers to be derived. One step involves obtaining features from the image through explicit or implicit feature engineering. In addition to analytical validation, clinical validation is critical to ensure a biomarker is ready for clinical translation. Whilst some highly-defined machine learning classification tasks, such as lung cancer screening to diagnose whether a nodule is benign or malignant, using highly-selected data appear promising, few, if any techniques are ready to be incorporated into the clinic for routine use. Almost all machine learning studies, particularly where there is a combination of a relative scarcity of data and the clinical question is complex e.g. brain tumour treatment response, would benefit from improvements to their methodology. Examples include the use of using external validation datasets, comparison of the novel approach to simpler standard approaches and the use of larger datasets. Because the development and validation of machine learning models require large, well-annotated datasets, multidisciplinary and multi-centre collaborations are typically necessary. Given the challenges of translating machine learning models to the clinic, some groups have published roadmaps and scoring systems to improve study design: examples include the Imaging Biomaker Roadmap, the Image Biomarker Standardisation Initiative and the Radiomics Quality Score.