Hybrid descriptive-inferential method for key feature selection in prostate cancer radiomics

Abstract

In healthcare industry 4.0, a big role is played by radiomics. Radiomics concerns the extraction and analysis of quantitative information not visible to the naked eye, even by expert operators, from biomedical images. Radiomics involves the management of digital images as data matrices, with the aim of extracting a number of morphological and predictive variables, named features, using automatic or semi-automatic methods. Multidisciplinary methods as machine learning and deep learning are fully involved in this field. However, the large number of features requires efficient and effective core methods for their selection, in order to avoid bias or misinterpretations problems. In this work, the authors propose a novel method for feature selection in radiomics. The proposed method is based on an original combination of descriptive and inferential statistics. Its validity is illustrated through a case study on prostate cancer analysis, conducted at the university hospital of Palermo, Italy.