Page 13 - AIH-1-4

Artificial Intelligence in Health Radiomics in early-stage lung cancer

together on segmentation, the accuracy of segmentation Furthermore, data must be distributed evenly, which
can be increased. otherwise would increase the risk of overfitting. Radiomics
Differences between centers in terms of obtaining studies are still ongoing at the clinical trial level, and there
imaging studies should be eliminated. Several parameters exist several parameters that require standardization. In
require standardization, such as the device used for this regard, clinicians and engineers must standardize these
imaging, use of contrast in CT, slice thicknesses, and the parameters through a joint investigation. Standardization
time between nuclear material injection and imaging. of parameters is essential for the use of radiomics in
routine clinical practice; otherwise, each study may
Extraction of features involves a comprehensive identify different radiomics features as important variables,
quantification of tumor phenotypes. After image acquisition, resulting in varying accuracy rates of different algorithms
preprocessing, and segmentation, radiomics features are in multiple studies. Considering this heterogeneity, it
extracted from two-dimensional (2D) or 3D ROIs in the would be inappropriate to use radiomics as a standard in
images. A standard method for selecting ROIs is also lacking. routine clinical practice.
For metastatic diseases, there is still no clear consensus on
whether the metastatic focus or the primary focus should As technology advances, the importance of data
be segmented. Moreover, only the tumor and tumor also increases. Performing tumor and organ-at-risk
microenvironment, which is usually created at a 5-mm segmentation in the radiotherapy planning stage is
margin to the tumor, can be used as ROIs for feature a routine procedure for radiation oncologists and is
extraction. Furthermore, there is no consensus on whether included in the treatment planning of each patient. From
radiomics features should be obtained from the original this perspective, algorithms need to be trained on a large
image or filtered images. 24 amount of processed data in radiation oncology clinics.
These valuable data should not be ignored and must be
There are different types of radiomics features, such considered to contribute significantly to standardization.
as shape, first-order, and textural features. As it remains
unclear which feature should be used at which stage of In summary, radiomics is a promising non-invasive
tumor, most studies investigate all radiomics features. method for prognostic prediction and post-treatment
There is also no consensus on whether 2D or 3D features follow-up, wherein medical images can be analyzed. It has
should be used. The 2D radiomics features have been the potential to facilitate personalized treatment selection
reported to be superior in some studies, whereas 3D with low cost and high sensitivity. If the appropriate
features are reportedly superior in other studies in terms personalized treatments are administered, patients can
of prognosis. 25,26 avoid unnecessary treatment and high treatment-related
costs. Radiomics can be used as an important biomarker
Due to the lack of knowledge on which type of
radiomics features should be used in which tumors, in treatment decisions, prognosis determination, and
follow-up after accurate standardization.
hundreds of radiomics features are extracted; however,
if all of them are used to create a model, it will generate 5. Delta radiomics
an excessive number of features, causing confusion and
overfitting of the data and decreasing the actual accuracy. Radiomics has been developed to evaluate feature changes
It is important to consider that the extracted radiomics at different time points, an approach often referred to as
features may be related to each other, and feature selection “delta radiomics.” This method is used to examine the
should be conducted before modeling. It is also necessary effects of feature changes after certain steps in the patient’s
to use radiomics feature selection methods that identify the workflow (for instance, after a certain treatment). 28-30
most important features and remove redundant ones. As a Delta radiomics is a promising field of research and allows
standardized method for determining important variables changing and modulating the treatment approach due to
has not yet been developed, different methods are used in its predictive power. 31
various studies. 27 The vast majority of radiomics methods used in existing
After determining the important variables among studies depend on imaging data obtained at a single time
100 of radiomics features using appropriate methods, an point, often imaging tumors before the start of treatment.
accurate model is established. Researchers use different Delta radiomics reveals the changes in feature values
machine learning algorithms to create models. Each study during treatment by extracting quantitative features from
has reported different algorithms with the best accuracy. image sets obtained throughout the treatment process. 32,33
This leads to further confusion among clinicians. Thus, Delta radiomics may improve cancer diagnosis, prognosis,
the most effective algorithm for determining prognosis patient follow-up, or evaluation of treatment response. 34,35
remains unclear. Some studies have shown that delta radiomics is effective

Volume 1 Issue 4 (2024) 7 doi: 10.36922/aih.3541