272 chapter 3 in, garbage out’. Furthermore, AI algorithms are often considered as black boxes, since they usually lack an easy and intuitive interpretation that can be interpreted in the domain of biology or radiology [376, 377]. Explainable AI (XAI) is developed to facilitate the interpretation of data in the context of a specific application and to retrace the results on demand [376]. Moreover, AI methodology is often heterogeneous and not unambiguously reported, complicating validation of the model. Model validation is a crucial step towards clinical translation, verifying whether the model is predictive for the general target population or just for a particular subset of patients. Models must be validated using an independent test set, preferably using data from a different institution. Currently, a lack of this external validation is still one of the major limitations of AI, while replication might be of even more scientific value than original discoveries [378]. Since 2012, AI analysis of a large number of quantitative variables derived from medical images has been studied in the field of radiomics [379]. Radiomics consist of the conversion of (parts of) medical images into a high-dimensional set of quantitative features and the subsequent mining of this dataset for potential information useful for the quantification or monitoring of tumour or disease characteristics in clinical practice. The field of radiomics includes the extraction of predefined, handcrafted intensity (i.e., first order), shape and texture features combined with statistical methods or machine learning algorithms for modelling; and more recent deep learning algorithms that both learn features from raw data and perform modelling (Figure 2) [380]. To create a holistic model, in addition to the imaging features, clinical characteristics or other -omics data, like genomics, proteomics or metabolomics, are also incorporated [381]. Radiomic analysis aims to find stable and clinically relevant image-derived biomarkers for tumour characterisation, prognostic stratification and response prediction, thereby contributing to precision medicine. In this chapter, the umbrella term radiomics encompasses a broad spectrum of image analysis methods, ranging from simple AI-based methods to sophisticated deep learning algorithms. The promises of radiomics were high. Hypothesising that medical images contained much more information than could be assessed by the human eye, radiomics was expected to contribute to medical decision making on a large scale and even to provide new insights in disease processes [362]. Yet, as for any new technology, many (technical and statistical) challenges have to be faced before reaching the goal of large-scaled implementation in clinical practice. Radiomic features are sensitive to technical variations in the different steps of the radiomic pipeline (Figure 2), hampering the reproducibility, validation and clinical translation of radiomic research. These technical variations should be as small as possible in order to attribute differences in feature values to tumour biology instead of technical variation. Image acquisition and reconstruction largely contribute to data inhomogeneity. Radiomic analysis often consists of retrospective analysis of standard-of-care images and reanalysis of previously published cohorts, where scanners and scan protocols may vary widely between different
RkJQdWJsaXNoZXIy MTk4NDMw