2. Radiomics and AI Workflow

In precision medicine, radiomics is currently underway in research based on feature extraction from medical images. Radiomics paves the way to map multimodal imaging into quantitative information on a large scale [24]. Figure 3 depicts the steps required to build a predictive model from medical images ^[25][26][25,26]. The radiomics workflow begins with image acquisition and segmentation. After image post-processing, hundreds of radiomic features (RFs) are measured from segmented regions to provide raw data for developing the final model. The major categories of features are as follows:

• Geometric or shape features: based on the segmented regions.
• Statistical or intensity features: computed using intensity values in each image region.
• Textural features (TFs): quantification of image intensity and regularity via mathematical functions.
• Wavelet or high-order features: the image transformation process is essential to obtain these features.

Individual features are discarded for dimension reduction through feature selection in the next step. Options include intraclass correlation coefficients (ICC), principal component analysis (PCA), least absolute shrinkage and selection operator (LASSO), recursive feature elimination (RFE), and outputs from machine learning methods (ML). Different approaches are applied in developing the model depending on the task: univariate or multivariate analysis and supervised or unsupervised ML methods. Supervised ML algorithms are classified into classification and regression algorithms if the variables are categorical or continuous, respectively. Classification algorithms are divided into linear and nonlinear models. Logistic regression and support vector machines (SVMs) are two methods of analyzing linear models. In nonlinear models, k-nearest neighbors (KNNs), gradient boosting, decision tree, extra trees (ETs), and random forest are the most commonly used algorithms. The most widely used regression algorithms are linear, logistic, polynomial, support vector regression (SVR), decision regression, random forest regression (RFR), and ridge and lasso regressions. In contrast to supervised learning methods, unsupervised learning approaches do not contain predefined response variables. Instead, the model finds hidden patterns and insights from the given data. In these procedures, similar data are grouped (clustering), or dimensionality is reduced. Some popular models in this category are K-means clustering, KNN, neural networks (NNs) or artificial neural networks (ANNs), PCA, and independent component analysis (ICA) [27]. Deep learning (DL) methods have been introduced as a more comprehensive part of ML methods with various techniques. Classic neural networks, convolutional neural networks (CNNs), recurrent neural networks, auto-encoders, generative adversarial networks (GANs), and gradient descent are examples of DL methods [28]. Challenges still need to be addressed to strengthen radiomics’ role in clinical practice. Most difficulties come from imaging feature variability among different devices and protocols, model robustness, and performance interpretation. In a multicenter context, addressing variability in acquisition and reconstruction protocols is crucial to ensure reproducibility [29]. Accordingly, harmonization procedures have been developed to provide a high reproducibility of RFs in multicenter studies. Reuze et al. [30] and Orlhac et al. [31] reviewed the radiomics workflow and its challenges. The Society of Nuclear Medicine and Molecular Imaging (SNMMI) AI Task Force published a report on evaluating and validating AI algorithms [32]. This guideline applies extensively to radiomics studies involving AI. According to Figure 4, created based on this guideline, for a prostate cancer patient referred to PET/CT imaging, AI can be applied in a chain from radiochemistry to the physician’s report generation. In the first step, AI could predict drug-target interactions, predict and optimize radiochemical reactions, carry out de novo drug design, and optimize radiopharmacy workflows. In the next step, ML-based methods may be well suited to difficult issues in image acquisition and instrumentation. For image reconstruction, AI may offer faster image reconstruction, a better signal-to-noise ratio, and fewer artifacts. Image analysis can be automated using AI for different tasks, such as lesion detection, segmentation, and quantification for diagnosis and dosimetry. Moreover, AI has the potential to investigate patterns associated with patient results within large biological and imaging datasets. Additionally, AI can also detect and diagnose. By using ML methods, diagnostic images can be interpreted and translated into reports and clinical databases. Finally, clinicians can receive actionable advice after extracting, distilling, and integrating clinical information from various sources.