Reproducibility and interpretability in radiomics: a critical assessment

Abstract

adiomics is a recent field that uses "an automated high-throughput extraction of large amounts of quantitative features of medical images." 1-3The method "converts imaging data into a high dimensional mineable feature space using a large number of automatically extracted data-characterization algorithms." 4The above definition may seem complex, but it can be succinctly summarized.Similar to how clinical routine involves characterizing a patient using parameters such as age, weight, and hemoglobin levels, radiological images can be analyzed to extract analogous parameters (also called features) that ideally describe the pathology of interest.For example, in the case of a tumor lesion, features such as its volume and diameter can be measured.A critical aspect of radiomics is the extraction of not only morphological features but also the distribution of intensity and texture.This includes, for instance, assessing whether the lesion has high brightness and a homogeneous or coarse texture, and identifying the presence of bright spots.Radiomics involves the extraction of hundreds to thousands of such features to accurately represent the lesion.These features are subsequently used to train a classifier, that, based on the characteristics of a new lesion, can determine, for example, whether the lesion is benign.The main expectation of radiomics is that these features can serve as surrogates for biomarkers, and thus aid clinical decision making.Radiological imaging could reflect the underlying biological processes, allowing for indirect conclusions.For example, while necrotic cells are not directly observable in computed tomography (CT) scans, their presence may result in the appearance of a hypodense lesion (Figure 1).Thus, measuring the overall intensity of a lesion could be used as an indicator of cell necrosis.Although radiomics as a field only emerged in the 2010s, 1,5 the idea can be traced back much further.In a seminal paper published in 1978, Harlow et al. 6 introduced concepts that are strikingly similar.Later, specifically in the 1990s, similar techniques were introduced as texture analysis.7This is no coincidence, since the underlying idea of applying machine learn-

Topics & Concepts