Page 127 - Read Online

Page 167 De Robertis et al. Art Int Surg 2023;3:166-79 https://dx.doi.org/10.20517/ais.2023.18

diagnostic accuracy and reproducibility. Moreover, radiological images contain a large amount of
information that is invisible even to an experienced human eye.

Radiomics is an advanced computational analysis of biomedical images, including ultrasound (US),
computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography
(PET), that aims to obtain an objective, detailed, and multidimensional characterization of biological tissues
through the extraction of numerical features by converting images into mineable data. Radiomics features
numerically describe and quantify the spatial distribution and relationships between the voxels that
compose a “black and white” image, which ultimately reflect the underlying physiopathology of the tissue
under study . Although the clinical role of radiologists in patients’ care remains a cornerstone of cancer
[1]
imaging, the addition of radiomics features to the visual assessment of an imaging examination may
improve the diagnostic performance of radiologists, the reproducibility of the results, and the patients’
outcomes by identifying adverse pathological features, predicting disease recurrence and survival, and
improving the evaluation of treatment response, which are extremely important for HBP surgery. Previous
reports revealed that radiomics is a promising tool to improve the non-invasive characterization and
[1]
preoperative staging of HBP neoplasms . Nevertheless, each individual step in the process of radiomics has
technical challenges that result in a significant translational gap between research and clinical practice.

The aim of this paper was to review the current role of radiomics in HBP surgery by analyzing systematic
reviews, meta-analyses and the most relevant published series.

RADIOMICS: WORKFLOW, POTENTIAL AND LIMITATIONS
The most important field of application of radiomics is oncological imaging. The underlying hypothesis is
that radiomics features parallelize the heterogeneity that characterizes tumor histology, allowing deep
exploration of tumor microenvironment and intra- and inter-tumoral heterogeneity, which are ultimately
[3]
[2]
related to the biological and genomic characteristics . The workflow of radiomics consists of several steps :
(1) acquisition of standardized, high-quality radiological images; (2) accurate segmentation of the tumor
mass with delineation of a volume of interest (VOI); (3) extraction of reproducible, non-redundant and
uncorrelated radiomics features; (4) integration of radiomics features with pathological and clinical data; (5)
construction of a database for data mining.

Each of these steps is a potential source of bias that may affect the quality, robustness and reproducibility of
the results. Given that a single voxel can influence the radiomics features, differences in the equipment, for
example, the magnetic field strength in MRI and the number of detectors in CT, and the image acquisition
protocols lead to relevant discrepancies between studies.

Segmentation, defined as the delineation of the tumor mass relative to adjacent structures, is one of the
most important sources of variability. Manual tumor segmentation is time-consuming and limited by
human capabilities, but also has the advantage of being controlled in real time by the human eye; on the
other hand, automatic tumor segmentation is fast and highly standardizable, but lacks the ability to
iteratively understand whether data are being acquired correctly; finally, semi-automatic tumor
segmentation seems to be the most effective method because it combines the advantages of computer
[3]
technology with the control of the human eye . Radiomic feature extraction is a poorly standardized
process, given the multitude of software applications that work differently to convert voxels to numerical
data; furthermore, post-processing of biomedical images can be done with several different modeling
algorithms. Both these aspects increase the heterogeneity between studies. Finally, large, shared databases
for data mining would be essential to validate the results of single studies by interrogating separate