Page 128 - Read Online

De Robertis et al. Art Int Surg 2023;3:166-79 https://dx.doi.org/10.20517/ais.2023.18 Page 168

populations of patients and to obtain study cohorts with sufficient size for statistical power.

Radiomic research initially used manual tumor segmentation and conventional statistical tests, such as the
Fisher’s or Wilcoxon tests, to select radiomics features; more complex statistical analyses, such as the least
absolute shrinkage and selection operator (LASSO), were then used to select predictive features. Although
this approach is easy to perform even by operators without specific experience in image processing and
statistical analysis, it is time-consuming and its application is limited to small populations. The current
research trend relies on machine learning (ML) methods, as they allow for a higher level of automation
compared to the traditional workflow, providing faster segmentation and feature extraction as well as
advantages in terms of reproducibility. Despite these benefits, ML-based radiomics are limited by the high
correlation to the quality of the input data (i.e., the accuracy of the segmentation and the size of the training
population); therefore, large datasets are necessary to identify robust features.

Researchers’ interest in radiomics has exploded in recent years, and thousands of studies on several different
settings have been published. As a decade of radiomics research is approaching, it is time for a critical
review of what results have been achieved and what has been translated into clinical practice. The
[4]
Radiomics Quality Score (RQS) was developed to evaluate the methodological quality of a radiomic study .
The RQS consists of 16 criteria, with a score of 36 (or 100%) indicating excellent study quality [Table 1]. A
[5]
recent study reported a median RQS of 21% among 44 systematic reviews on radiomics, suggesting that
the quality in this research field needs to be increased as it is currently unsatisfactory regardless of the topic.

METHODS
A search of the MEDLINE database was performed to identify meta-analysis and systematic reviews
relevant to detection, characterization and differential diagnosis, identification of adverse pathological
features, and prediction of prognosis of hepato-bilio-pancreatic neoplasms using radiomics. The following
terms were searched: “(liver OR hepatic) AND (radiomics OR histogram analysis)”, “pancreas AND
(radiomics OR histogram analysis)”, “(hepatocellular carcinoma OR hepatocarcinoma) AND (radiomics
OR histogram analysis)”, “cholangiocarcinoma AND (radiomics OR histogram analysis)”, “liver metastases
AND (radiomics OR histogram analysis)”, “pancreatic adenocarcinoma AND (radiomics OR histogram
analysis)”, “pancreatic neuroendocrine AND (radiomics OR histogram analysis)”, “intraductal papillary
mucinous neoplasm AND (radiomics OR histogram analysis)”. A total of 38 meta-analyses and systematic
reviews were retrieved; the search was further expanded by reviewing the reference list of the selected
articles in order to identify relevant papers in terms of sample size, originality, methodology and importance
of the results. Full-text articles in English, reporting on US, CT, and MRI, published by 15th May 2023, were
considered. Studies were excluded if they were not conducted in humans or if they evaluated other imaging
techniques such as PET-CT or endoscopic US. Seventy-one studies met the eligibility criteria and were
included in this literature review. The results of the meta-analysis and systematic reviews that reported
pooled data diagnostic values are summarized in Table 2; the studies analyzed are described in the following
paragraphs.

HEPATOCELLULAR CARCINOMA
Hepatocellular carcinoma (HCC) is usually diagnosed and monitored by imaging; therefore, it is a good
candidate for radiomic analysis. Several studies have been published regarding the identification, diagnosis,
and prediction of adverse pathological features, and prognosis in HCC patients.