Abstract

Although data-driven Artificial Intelligence (AI) methods have played a crucial role in recent years, their actual implementation in medicine still poses numerous challenges. The lack of transparency has fostered skepticism among physicians and patients towards these emerging technologies. Notably, the US Federal Trade Commission emphasizes the importance of transparency, explainability, fairness, and other ethical considerations in the use of AI tools. Additionally, under the GDPR , individuals have the right to receive an explanation regarding the predictions made by AI systems. In response to these concerns, several post-hoc explanation methods have been proposed to elucidate the deep features extracted by neural networks, including Convolutional Neural Networks, Graph Neural Networks, and Vision Transformers. However, these methods have exhibited limitations in effectively correlating predictions with clinical justifications. Recently, particularly in the field of radiological medical imaging, Radiomics has emerged as a powerful tool for feature extraction. Unlike deep features, radiomic features are derived through mathematical formulas applied to the images, enabling the extraction and quantification of various image characteristics such as texture, pattern, and statistical measurements of regions of interest. Radiomic workflows offer several advantages over deep feature extraction, including the ability to work with small datasets (a common scenario in the medical field). Moreover, the inherent interpretability of radiomic features is well-established, as each feature carries a known meaning. The combination of Radiomics and explainable-by-design methods holds great promise in advancing the transparency and interpretability of AI applications in medical imaging.