Radiomics in medical imaging: A brief introduction

The aim of the present blog is to give a brief introduction to radiomics and present a typical workflow of radiomics analysis.

Image credit: Review on Radiomics for Brain Tumors.

Key points

• Radiomics is a technique for the quantitative description of multi-modal medical images.
• An overview on radiomics analysis typical workflow.
• Useful links and and open-source resources that can help enhance your understanding of radiomics.

AI in medicine

Photo credit: Roadmap for AI in Medical Imaging.

In recent decades, the medical industry has experienced a consistent rise in the digitization of clinical data. Additionally, the methods of artificial intelligence (AI) have become not powerful and accessible enough to put the focus on identifying the most relevant data and the information most likely to be helpful to solve a given issue. Consequently, there has been a substantial increase in AI applications in healthcare, with the aim of simplifying repetitive tasks clinicians encounter in their daily workflows and promoting the advancement of personalized medicine.

Machine learning (ML) is a subset of AI with promising tools that can assist the medical field in tasks like classification, segmentation, and more. Deep learning (DL) is a specific branch of ML that involves using multiple layers of neural networks to detect and/or extract features in imagery. Most AI-based medical field tasks are conducted using DL algorithms. However, in the realm of ML-based classification of medical imaging data, radiomics is a well-established competitor to DL. When compared to DL, radiomics features are highly explainable since they are based on mathematical definitions. As a result, radiomics is highly capable of handling massive amounts of data in the medical field, with the goal of delivering precision medicine.

What is radiomics?

Radiomics refers to the high-throughput quantitative image analysis which aims at extracting a large number of features characterizing a region of interest in the medical image. Radiomics analysis can be performed on medical images from different modalities, e.g., from from magnetic resonance imaging (MRI), computed tomography (CT), and positron-emission-tomography (PET). This approach enhances the existing data available to clinicians by quantifying the spatial distribution of signal intensities, voxels interrelationships and textural information.

Radiomics typical workflow

The figure below demonstrates the necessary stages in a typical radiomics pipeline, emphasizing the crucial aspects of the radiomics analysis procedure.

Radiomics analysis typical workflow

• Segmentation: For every radiomics analysis, the region of interest (ROI) definition is crucial for the extraction of information from medical images. The images segmentation might be done manually by en expert or automatically using a deep learning model, such as MONAI’s model Swin UNETR.

Brain tumors segmentation. Image credit: paperswithcode.

• Image processing: Image processing comes after the segmentation step and before the feature extraction step. It aims to standardize images, from which radiomic features will be extracted, in terms of pixel spacing, gray-level intensities, histogram bins, etc. The main processing steps are: Interpolation, re-segmentation, ROI extraction and intensity discretization. Note that the image processing steps leading to features computation can be carried out in various ways, in terms of processing parameters. A separate blog post will be written specifically for the image processing steps.

ROI extraction using an ROI mask. Intensities around the ROI are excluded. Image credit: IBSI documentation figure 6.

• Features extraction: Following image segmentation and processing, features can be extracted from the image. Feature extraction is the final processing step where feature descriptors are utilized to quantify characteristics of the grey levels within the ROI. The set of features can be classified into various families, including intensity-based statistical features, intensity histogram-based features, intensity-volume histogram-based features, morphological features, local intensity features, and texture matrix-based features. Each family has the potential to reveal tumoral patterns, size, shape, and textural characteristics that may not be discernible to the naked eye. Additionally, features can be computed on the base image, as well as from filtered images.

Feature categories

Radiomics extraction packages and platforms

Numerous open-source software platforms and packages are available to extract radiomics features from medical images. For instance, PyRadiomics, an open-source Python-based package that aims to establish a reference standard for radiomic analysis. In addition, LifEx is a standalone platform that integrates segmentation and texture analysis tools with a user-friendly graphical interface. Our laboratory, MEDomics-Udes, has developed an open-source Python package for radiomics extraction named MEDimage. This package offers various modules that can be utilized for radiomic studies, as well as additional functionalities that assist users in processing their datasets and selecting the settings for their studies. MEDimage is available on GitHub and is continually improving.

Since many different ways and formulas exist to calculate those features, adherence to the Image Biomarker Standardization Initiative (IBSI) guidelines is recommended. These guidelines offer a consensus for standardized feature calculations from all radiomic features.

Useful links for radiomics

• The MEDimage application: https://medomics-udes.gitbook.io/medimage-app-docs
• The MEDimage package: https://github.com/MEDomics-Udes/MEDimage.
• Radiomics in medical imaging — “how-to” guide and critical reflection: https://doi.org/10.1186/s13244-020-00887-2.
• IBSI: https://theibsi.github.io/.