Page 152 - Read Online

Grewal et al. Art Int Surg 2023;3:217-32 https://dx.doi.org/10.20517/ais.2023.28 Page 219

[18]
to therapy and risk of recurrence, thereby improving survival . This narrative review focuses on the
current literature on radiomics as a biomarker for hepatobiliary and pancreatic malignancies. We discuss
how radiomics could help guide the management of these diseases, the current limitations of radiomics, and
future applications and integration in the clinical setting.

METHODS
To identify and synthesize literature for a narrative review regarding the utility and applications of
radiomics in the management of HPB tumors, PubMed and Embase, as well as Google Scholar, were
queried. These platforms were searched from inception until July 2023. Given that a qualitative review was
planned, all article types published in English were considered eligible for inclusion. Terms such as
“radiomics” and “segmentation” were combined with the term, “biomarker”, and various iterations of
aspects of management, such as “prognostication” and “treatment response” and specific HPB tumors (e.g.,
“HCC” “PDAC”, etc.).

DEFINING RADIOMICS
Recently, the field of radiomics has burgeoned and shown promise as a potential tool for early diagnosis,
tumor characterization, and prognostication . This involves extraction of high-dimensional data from
[19]
images and providing feature data for quantitative description of lesions . Radiographic images contain a
[20]
[19]
number of quantifiable features that may be mined and analyzed to offer insights into disease processes .
The general radiomics workflow can be broadly summarized into four main steps: image acquisition, region
[21]
of interest (ROI) segmentation, feature extraction, and analysis [Figure 1] . Image segmentation consists of
the delineation of the boundaries of a ROI, such as a tumor and adjacent anatomical structures. This step
can be done manually, semi-automatically, or fully automatically with deep learning algorithms. Manual
segmentation and semi-automatic segmentation, though time-consuming, have been the most encountered
[22]
methods of segmentation in current literature . While these methods of segmentation are prone to
introducing inter-observer biases stemming from inconsistencies in the delineation of the boundaries of
ROIs, prior studies have uncovered conflicting results on the influence this actually has on radiomics
features [21,23] . Once ROIs have been segmented, the radiomic features are extracted via the conversion of
imaging data into quantifiable features, such as signal intensity, shape, texture, and higher-order features.
Signal intensity features are obtained through analysis of histograms of individual voxel signal intensities
and offer insights into the central tendency of the distribution. Shape and texture features are both
calculated in three dimensions by considering the correlation of signal intensities of surrounding voxels.
Higher-order features may also be extracted following the application of secondary wavelet or Gaussian
[22]
filters . The number of features extracted through these processes can significantly vary depending on
filter and software specifications. Overfitting of the model may occur in cases where a high number of
features are coupled with a low number of cases in a classification task. The next step of the workflow,
feature selection, mitigates this risk by selecting relevant features using techniques such as random forest
algorithms and by excluding non-reproducible features. Relevant features may then undergo subsequent
analysis using machine learning algorithms.

APPLICATION OF RADIOMICS IN PANCREATIC CANCER
Pancreatic cancer is one of the leading causes of cancer-related deaths globally, with an estimated 5-year
survival of approximately 12% . As compared to other malignancies, the incidence of PC is on the rise
[24]
while the outcomes remain poor. Over the last four decades, minimal improvement in survival has been
observed, the only major development being the introduction of multiagent systemic therapies in the last
decade . The two key challenges that are faced in the management of these patients are the lack of
[25]
screening tools, resulting in a delay in diagnosis and limited biomarkers that can help guide management in