In the field of medicine, radiomics is a method that extracts a large number of features from medical images using data-characterisation algorithms. These features, termed radiomic features, have the potential to uncover tumoral patterns and characteristics that fail to be perceived by the naked eye. The hypothesis of radiomics is that the distinctive imaging features between disease forms may be useful for predicting prognosis and therapeutic response for various cancer types, thus providing valuable information for personalized therapy. Radiomics emerged from the medical fields of radiology and oncology and is the most advanced in applications within these fields. However, the technique can be applied to any medical study where a pathological process can be imaged.
Process
Image acquisition The image data is provided by radiological modalities as CT, MRI, PET/CT or even PET/MR. The produced raw data volumes are used to find different pixel/voxel characteristics through extraction tools. The extracted features are saved in large databases where clinics have access so as to enable broadly collaborative and cumulative work in which all can benefit from growing amounts of data, ideally enabling a more precise workflow.
Image segmentation After the images have been saved in the database, they have to be reduced to the essential parts, in this case the tumors, which are called "volumes of interest". Because of the large image data that needs to be processed, it would be too much work to perform the segmentation manually for every single image if a radiomics database with lots of data is created. This makes manual contouring time-consuming and can be potentially affected by rater variability. Instead of manual segmentation, an automated process has to be used. Two possible solutions are deep-learning-based automatic tumor segmentation methods and semiautomatic segmentation workflow where human label the first few slices and then the ML system do the rest. Before it can be applied on a big scale, an algorithm must score as high as possible in the following four tasks:
First, it must be reproducible, which means that when it is used on the same data the outcome will not change. This can either be different raters’ contour is similar or the automatic/semiautomatic models’ have good robustness measured by Intraclass Correlation Coefficient. Another important factor is consistency. The algorithm does solve the problem at hand and performs the task rather than doing something that is not important. In this case, it is necessary that the algorithm can detect the diseased part in different scans. The algorithm also needs to be accurate. It is very important that the algorithm detects the diseased part in the most precise way possible. Only with accurate data, accurate results can be achieved. Besides approximate visual effect, typical metrices Dice similarity coefficient, Jaccard index, Hausdorff distance, and average surface distance, which gives more quantitative and objective evaluation. A minor but still important point is time efficiency. The results should be generated as fast as possible so that the whole process of radiomics can also be accelerated. Usually automatic/semiautomatic segmentation algorithms are more efficient than manual segmentation. For automatic/semiautomatic segmentation, training time and running speed are usually the two most important standards. To speed up both standards, typical solution is structure / architecture-level change, like knowledge distillation-based framework. In recent years, machine learning and deep learning have been the dominant part of automatic medical image segmentation. Many methods are based on convolutional neural networks, such as U-Net-like models and CE-net.
Feature extraction and qualification After the segmentation, many features can be extracted and the relative net change from longitudinal images (delta-radiomics) can be computed. Radiomic features can be divided into five groups: size and shape based–features, descriptors of the image intensity histogram, descriptors of the relationships between image voxels (e.g. gray-level co-occurrence matrix (GLCM), run length matrix (RLM), size zone matrix (SZM), and neighborhood gray tone difference matrix (NGTDM) derived textures), textures extracted from filtered images, and fractal features. The mathematical definitions of these features are independent of imaging modality and can be found in the literature. A detailed description of texture features for radiomics can be found in Parekh et al. (2016) and Depeursinge et al. (2017). Due to its massive variety, feature reductions need to be implemented to eliminate redundant information. Hundreds of different features need to be evaluated with a selection algorithm to accelerate this process. Additionally, features that are unstable and non-reproducible should be eliminated since features with low-fidelity will likely lead to spurious findings and unrepeatable models.
Analysis Analysis After the most relevant features are selected, the data must be analyzed to find meaningful patterns and relationships. In modern medical analysis, AI, machine learning, and deep learning are often used to analyze imaging and non-imaging patient data for tasks such as disease detection, lesion characterization and treatment planning. This information can then be further used in forming CAD (computer aided diagnosis) systems. In radiomics, low order image features such as shape, intensity, texture, and then use traditional machine learning methods such as support vector machines, random forests, logistic regression, PCA, and clustering to make predictions, or deep learning methods such as convolutional neural networks that learn features automatically from the data in convolutional layers and then make predictions or regression on the fully connected layer. Before the actual analysis, different data types such as imaging radiomics, molecular biomarkers, and genetic information can be integrated to enhance model’s performance. This multiomics integration allows AI models to make decisions based on multi-dimensional data, which can improve cancer subtyping, risk prediction, treatment selection, and clinical decision support. The integrated data can then be analyzed using supervised learning, where uses an outcome variable to be able to create prediction models, or unsupervised learning to summarize the information we have and can be represented graphically.
Databases
… excerpt ends here. Continue reading the full article.
