The world’s leading publication for data science, AI, and ML professionals.

What is Radiomics?


Radiomics broadly defines a process that involves extracting high-throughput quantitative features from radiographic images.

Radiomics stems from decades of medical imaging and computer-aided diagnosis. In this field, medical images are converted into high-throughput quantitative features ( Lambin et al., 2012). These features have displayed the potential to enhances diagnostics and prognostic predictions (Rao et al., 2012). Although quantify features derived from medical images date back a few decades (Haralick., 1973), the ‘omics’ suffix expresses the dimensionality of this data, the biological importance, and connects these features with precision medicine.

A radiomics study comprises five main phases: participant selection, imaging by clinicians, radiomics feature extraction, exploratory analysis, and modelling. Lambin and colleagues have proposed a radiomics quality score, a 36 point system to increase the robustness of the entire radiomics pipeline (Lambin et al 2017).

Although radiomics is a novel definition of high throughput feature extraction from medical images, there are two types of radiomics feature extraction methods; "Conventional radiomics" and ‘Deep radiomics’.

Figure1: The evolution of radiomics. (Image by author)
Figure1: The evolution of radiomics. (Image by author)

Deep Radiomics

Deep Learning workflows can be used to learn features from images and perform classification/regression without the need for detailed ROI delineation.

Fundamental to this approach is that deep learning models should not only be used for data mining and predictions but also for data generation. The number of features extracted using deep learning is several times larger than the handcrafted feature. Deep radiomics has several advantages including, being capable of generated vast numbers of features and being able to classify within the same pipeline.

Whilst deep Radiomics offers potentially endless numbers of features, there are three main challenges. The first challenge is feature interpretability. Although we can associate deep radiomics features with biological features (genomics, transcriptomics etc.) Understanding exactly what each radiomcs feature means is a challenge, especially when compared to handcrafted radiomics features. However, with the modern advancements in explainable deep learning, by using techniques such as activation mapping, there is scope for deep radiomics to diminish this lack of interpretability.

The second challenge is over-fitting. Over-fitting is a problem when the model fits very well on the training data set but not on the test/ validation data sets. This is often a consequence of a large number of features (thousands to hundreds of thousands) per a relatively low number of samples.

Whilst techniques such as transfer learning aid the use of deep learning on small data sets. A third issue is based on the use of transfer learning in deep radiomics. The overarching idea with transfer learning is to use a pre-trained model, that has been training on hundreds of thousands to millions of images to recognise important features. Without transfer learning, training on a small cohort would not learn these important features. However, the majority of pre-trained models are trained on 2-Dimensional images and do not apply to the raw 3D medical image.

Conventional Radiomics Features

Conventional Radiomics is the method whereby regions of interest (ROI) are delineated and ‘handcrafted’ features are extracted. These features attempt to describe the ROIs by capturing parameters that describe ROI morphology, shape and texture (Wang et al., 2014). These features are then associated with outcome variables, such as survival time.

Shape features are as the term sounds, they describe the shape of the delineated ROI, and properties including its volume, maximum diameter, maximum surface, tumour compactness, and sphericity.

First-order statistics features are essentially summary statistics of the voxel values. This includes, mean, median, minimum, maximum, skewness, kurtosis, uniformity, and entropy.

Second-order features, also known as textural features are obtained from voxel to neighbouring voxel relationships. Textural features provide information on voxel intensities and geometry. A common set of textural features are those derived from the grey level co-occurrence matrix, which quantifies the number of voxels with the same intensities within a given range. Another example is textural features derived from the Grey-level run-length matrix (GLRLM), which counts the number of consecutive voxels with the same intensity. The key difference between the GLCM & GLRLM being the count of consecutive voxel intensities.

Higher-order statistics features are often first or second-order features, derived from images that have been mathematically transformed or filtered. The idea is they will suppress unwanted variation/noise or reveal important details/patterns. Techniques used to calculate these features include fractal analysis, functional analysis with Minkowski functional, wavelet transformations, and Gaussian Laplacian transformations.

With the increasing number of conventional radiomics features based techniques, software, and their complexity, the Image Bio-marker Standardisation Initiative (IBSI) has outlined a series of feature extraction guidelines and feature nomenclature (Zwanenburg et al., 2016).

Radio-genomics

The term "radio-genomics" was initially used to describe the radiotherapy-induced toxicity on the genetic profile of a tumour (Kang et al., 2018). However, recently radio-genomics has been used to describe radiomics features and their associations to genomics and beyond, such as proteomics and metabolomics/metabonomics (Peeken et al., 2018). Although, using the term radiogenomics to describe a radiomics to metabolomics/metabonomics relationship is likely to add unnecessary confusion due to the genomics suffix.

Radiomics and Radio-genomics Challenges

Although the field of radiomics and radio-genomics holds potential, several challenges must be addressed before it will reach the clinic. (Shaikh., 2017).

Every step in the radiomics pipeline, from participant selection to radiomics modelling, is error-prone. This can lead to the propagation of error. For example, ROI delineations, have a degree of inter-observer variability. This is difficult to control, especially if ROIs have not been delineated by multiple radiologists and similarity score calculated, this may impact the downstream radiomics features (Leijenaar et al., 2013).

A study by Leijenaar and colleagues displayed for most radiomics feature the inter-class correlation coefficients (ICC) were greater than 0.9 between independent ROIs delineated by 5 observers, in PET-CT scans of a non-small lung cancer carcinoma cohort. Interestingly in Leijinaar’s study, the test-retest cohort had a lower ICC for the majority radiomics features (0.71 avg). However, this cohort was only in 11 patients compared to nearly 23 in the inter-observer cohort.

Collectives such as the IBSI, and Quantitative Imaging Biomarker Alliance are beginning to produce standard guidelines for the entire Radiomics pipeline, from patient scans to modelling, to increase the reproducible. To account for issues within the radiomics pipeline. Lambin and colleagues have proposed a radiomics quality score, a 36 point system to increase the robustness of the entire radiomics pipeline. The basis of this point system is 16 key elements. These involve questions such as if a phantom image was collected if multiple segmentation by different clinicians has been performed, and if works are open source.

Summary

Radiomics is a cheap, efficient, and non-invasive way to extract more information from medical images. Both conventional radiomics and deep radiomics yield several challenges. However, there is a potential for a huge medical revolution when these strategies become clinically viable.


Related Articles