Page 10 - AIH-1-4

Artificial Intelligence in Health Radiomics in early-stage lung cancer

significantly associated with both LC (>0.043 vs. ≤0.043, helpful in clinical decision-making. The abovementioned
15
P = 0.042) and MFS (>0.304 vs. ≤0.304, P < 0.001). studies are summarized in Table 1.
Considering clinical factors, their study suggested that The field of radiomics, which quantitatively analyzes
tumor histology, tumor diameter, and radiomics score are extensive data obtained from medical images, has significant
vital factors in predicting NSCLC recurrence patterns after potential for predicting treatment outcomes. Nevertheless,
SBRT. 13
a standardized imaging method has not yet been developed.
Lafata et al. evaluated the FB planning CT-based Moreover, the segmentation of the target region for the
radiomics features of 70 patients with early-stage collected data varies among researchers. Studies have been
NSCLC who received SBRT with a mean of 51 Gy at conducted using several varying radiomics features using
Duke University. Only two of 43 radiomics features, different software, and no consensus has been achieved
namely, Homogeneity2 and Long-Run-High-Gray-Level- on this issue. Multicenter radiomics-based prediction
Emphasis, were considered significant variables and were algorithms with higher accuracy can be generated using
associated with LC. The area under the curve (AUC) values a standard imaging method and artificial intelligence
of the multivariable LR prediction models for recurrence, (AI)-based segmentation.
local recurrence, and non-local recurrence were 0.72 ± 0.04,
0.83 ± 0.03, and 0.60 ± 0.04, respectively. These data suggest 3.2. PET- and CT-based radiomics models
that relatively dense tumors with homogeneous rough FDG PET-CT is a molecular imaging technique that
tissue are associated with higher rates of local recurrence. combines metabolic and functional assessments, improving
This was supported by both univariate and multivariate the diagnostic accuracy and initial staging and restaging of
analyses that showed that CT-based radiomics features lung cancer and influencing treatment optimization and
may be associated with local recurrences after SBRT in monitoring of treatment response. Although most studies
patients with Stage 1 NSCLC. 14 on PET-CT have shown that the standardized uptake value
Unlike other studies, Li et al. evaluated radiomics can be used as a prognostic indicator of OS, some studies
features in follow-up CT after SBRT. In their study, a dose do not support this. 16,17
of 50 Gy was administered in five fractions to 54 patients Evaluation of tumor heterogeneity is gaining importance
with early-stage NSCLC, 48 Gy was administered in four due to radiomics features obtained from different imaging
fractions to three patients, and 60 Gy was administered modalities. If prognosis can be predicted accurately
in eight or five fractions to two patients. The first control before treatment through radiomics, the most beneficial
CT was taken 1 – 3 months after the completion of SBRT personalized treatment can be administered to patients.
(median 91 days, range: 33 – 112 days) and used to extract Prognostic prediction studies have been conducted using
radiomics features. As imaging features, 34 manually a combination of radiomics and more classical PET
determined radiological features (semantics) describing parameters, and current prognostic prediction studies are
the lesion, lung, and thorax, and 219 quantitative imaging conducted using both PET and CT radiomics features to
features (radiomics) for the lesion were extracted. Cox create models with higher accuracy. 18,19
proportional hazards models and Harrell’s C index were
used to predict OS, relapse-free survival (RFS), and Dissaux et al. conducted a prognostic prediction study
locoregional RFS (LR-RFS), and a five-fold cross-validation using 18F-FDG PET-CT radiomics features in patients
was performed for the prognostic model. Data from a with early-stage lung cancer receiving SBRT. They included
median follow-up of 42 months were available. Eastern 64 patients from three different centers in the training
Cooperative Oncology for OS model Group performance set and 23 patients from one center in the test set. The
status, vascular involvement, lymphadenopathy, and primary tumor was segmented semiautomatically using
radiomics features considered important were included the fuzzy locally adaptive Bayesian algorithm on PET
in the study. Vascular involvement, pleural retraction, images and manually on low-dose CT images. ComBat
lymphadenopathy, vessel attachment, and relative was used to harmonize the radiomics features obtained
enhancement were included in the RFS model. In the from the four institutions, yielding 184 (92 PET and 92
LR-RFS model, vascular involvement, lymphadenopathy, CT) radiomics features. In the training set, significant
circularity, and radiomics features were included in variables in the univariate analysis were added to the
the study. The AUC, which was used to evaluate the multivariate regression model. Models were constructed
performance of the models, was >0.8 for all models. by combining the obtained independent prognostic
According to their study, the disease progression can factors. Two important variables were each obtained from
be predicted even 3 months after SBRT using the model PET and CT. The median follow-up duration was 21.1
established using CT imaging features, which might be (range: 1.7 – 63.4) and 25.5 (range: 7.7 – 57.8) months in

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