RADIOMICS REFERS TO THE AUTOMATED QUANTIFICATION OF THE RADIOGRAPHIC PHENOTYPE
Radiomic data has the potential to uncover disease characteristics that fail to be appreciated by the naked eye. The central hypothesis of radiomics is that distinctive imaging algorithms quantify the state of diseases, and thereby provide valuable information for personalized medicine. Radiomics has emerged from oncology, but can be applied to other medical problems where a disease is imaged.
Engineered features are hard-coded
features which are often based on expert domain knowledge. Measures include intensity, shape, texture, wavelet, and LOG features, and have been found useful in several clinical areas, such as oncology and cardiology.
Deep learning methods can learn feature representations automatically from data. These features are included in neural nets’ hidden layers. Within radiomics, deep learning involves utilizing convolutional neural nets - or convnets - for building predictive or prognostic non-invasive biomarkers.
Scientific studies have assessed the clinical relevance of radiomic features in multiple independent cohorts consisting of lung and head-and-neck cancer patients.