Page 40 - ARNM-3-2

Advances in Radiotherapy
& Nuclear Medicine Radiomics for gastric cancer

human epidermal growth factor 2, a cell-derived oncogene, To predict the survival risk of patients with GC, Zhang
as a bridge to predict the gene expression levels of patients et al. developed three models based on radiomics,
80
through radiomics, thus forecasting the prognosis of the clinical parameters, and deep learning, respectively, and
82
disease and therapeutic effect on cancer. they achieved good risk assessment results. Dong et al.
recruited 730 locally advanced GC patients from six centers
Studies demonstrated that the progression, metastasis,
and therapeutic response of cancer are determined by the and established a deep learning radiomic nomogram based
on multi-phase CT to determine the number of LNM in
tumor immune microenvironment. On the other hand, locally advanced GC patients before surgery. It was found
74
quantitative radiomic features extracted from images that the results are significantly better than those in the
are closely associated with molecular phenotypes. Two single algorithm models (C-index: 0.821 vs. 0.785).
13
82
studies applied non-invasive CT-based radiomics to Zhang et al. developed a nomogram combining radiomic
16
evaluate the tumor immune and immunosuppressive features, clinical parameters, and deep learning to predict
microenvironment to predict the treatment responses early recurrence in 669 multicenter advanced GC patients.
and outcomes for GC patients. 75,76 Similarly, Gao et al. The model achieved an AUC of 0.831, 0.826, and 0.806,
77
evaluated the abundance of tumor-infiltrating regulatory T respectively, in the training set and the two test sets, with
cells using CT radiomics to predict the outcome of GC. One the predicted risk showing good agreement with the
study measured the tumor immune microenvironment observed probability of recurrence.
based on PET-CT radiomics to predict the clinical
outcomes and chemotherapy response for patients with 4. Discussion and prospect
GC. 78 From the collected studies, radiomics has shown great
3.5. Deep learning radiomics potential in the diagnosis, clinical staging, and prognostic
prediction for patients with GC. In the differential
Machine and deep learning algorithms can mine vast diagnosis of GC, the evaluation effect of various radiomic
amounts of available image data to reveal underlying models was relatively stable (AUC = 0.770 – 0.903). 20-23
complex biological mechanisms and enable personalized However, in histopathological classification prediction, a
and precise cancer diagnosis and treatment plans. significant difference between the earlier model and the
79
Some studies have found that there are similarities and most recent model (AUC: 0.675 vs. 0.918) was observed
compatibility between radiomics and deep learning, and due to the addition of clinical features to the radiomic
the combined application of these two methods further features. 26-32 Pre-operative prediction of LNM is one of the
improves the prediction accuracy of GC, 16,80-82 as shown in most popular research directions. The prediction accuracy
Table 4. of the radiomic models obtained in most studies was more

Table 4. Summary of studies on radiomics combined with deep learning in gastric cancer
Authors Year Image Purpose Sample Features Result
Zhang 2019 CT To predict the early 669 (Multi-center) Radiomics+clinical+deep The nomogram, combining all these
et al. 16 recovery before surgery learning predictors, showed powerful prognostic
in patients with advanced ability in the training set and two test sets
GC with AUCs of 0.831 and 0.806, respectively
Zhang et al. 80 2021 CT To predict the risk of OS 640 (Multi-center) Radiomics+clinical+deep Deep learning model had the best
learning performance for risk prediction of OS
according to the C-index (training=0.82,
external validation=0.78)
Tan et al. 81 2021 CT To design a 86 (Single center) Radiomics+deep learning The predictive ability of the semi-automatic
semi-automatic segmentation was better than the manual
segmentation method segmentation (AUC of 0.828 and 0.749,
based on deep learning respectively)
Dong et al. 82 2020 CT Pre-operative LNM rate 730 (Multi-center) Radiomics+clinical+deep The model showed good discrimination
can be accurately assessed learning of the number of LNM in all cohorts
(C-indexes=0.821 in the primary cohort,
0.797 in the external validation cohorts, and
0.822 in the international validation cohort)
Abbreviations: AUC: Area under the curve; CT: Computed tomography; GC: Gastric cancer; LNM: Lymph node metastasis; OS: Overall survival.

Volume 3 Issue 2 (2025) 32 doi: 10.36922/arnm.8350