Page 9 - AIH-1-4

Artificial Intelligence in Health Radiomics in early-stage lung cancer

BED was 151.2 (range: 100 – 151.2) Gy. Radiomics features Sawayangi et al. randomly divided 358 patients treated
of all patients were obtained using both FB CT and AIP with SBRT into training (250 patients) and validation
CT. A total of 644 radiomics features were obtained (108 patients) groups and estimated the overall survival
using the MATLAB 2013 toolbox. Both FB CT and AIP (OS) using both clinical variables and CT-based radiomics
CT yielded 19 significant radiomics features, with six features. They applied 42 – 64 Gy SBRT in 4 – 10 fractions
features (two features of tumor shape, three features of (BED10 range: 75 – 166 Gy). To extract radiomics features,
intensity histogram or statistics of the tumor, and one the gross tumor volume (GTV) was contoured in the
feature of homogeneity of the tumor tissue) being similar expiratory phase of the planning CT. Radiomics features
in both image sets and the remaining 13 being completely categorized as gray level size region matrix features
different from each other but similar to each image type in calculated from GTV in the pretreatment CT image had
both image sets. The unique radiomics features of FB CT high accuracy for OS prediction in patients with early-
images were statistical (n = 6) and textural (n = 7) features. stage NSCLC treated with curative SBRT. These findings
Of the 13 unique radiomics features of AIP CT images, indicate that the operating system predicted from multiple
there were eight textural, four statistical, and one shape linear regression analysis is similar to the actual operating
feature. The median follow-up duration was 20.8 months. system. These data may be important in selecting patients
Distant metastasis (DM) developed in 20.5% of the who would benefit from increasing treatment intensity,
patients (n = 23), and the average time to metastasis was such as increasing the radiotherapy dose or using different
10.0 months. Locoregional recurrence (LRR) developed in fractionations and adding other treatments such as
21.4% of the patients (n = 24), and the median time until chemotherapy/immunotherapy/targeted therapies. 11
LRR developed was 8.8 months. The 2-year estimates for Luo et al. evaluated 119 patients and 129 lesions treated
LRR and DM were 70.9% and 74.0%, respectively. The with SBRT with a median of 48 Gy (range: 18 – 70 Gy) in a
concordance index (CI) was used to evaluate prognostic median of four fractions (range: 1 – 12). Like other studies,
performance. We found that FB CT radiomics features were they created clinical, radiomics, and combined models.
unrelated to DM development, whereas AIP CT radiomics They generated radiomics features based on planning
features describing tumor shape and heterogeneity were CT. Of 1502 radiomics features, four were identified as
compatible with DM development (CI: 0.638 ± 0.676). No important variables using the LASSO method. Logistic
imaging radiomics features were associated with LRR in regression (LR), decision tree, and support vector machine
their study. Compared with FB CT images, AIP CT images (SVM) algorithms were used to create the optimal model for
contained more valuable information regarding disease evaluating local tumor control. LR was determined as the
recurrence in patients with early-stage NSCLC receiving prediction algorithm with the highest accuracy rate, with
SBRT. According to their study, AIP CT images may be a the accuracy rates of the radiomics, clinical, and combined
valuable non-invasive technique to predict recurrence. 9 models being 67.4%, 82.0%, and 85.4% in the training
Kakino et al. conducted a local and distant recurrence group and 92.9%, 77.5%, and 82.5% in the validation group,
prediction study using breath-hold CT-based radiomics respectively. The combined model performed statistically
features in 573 patients with early-stage NSCLC who significantly better than the radiomics (P = 0.025) and
received SBRT. The patients were categorized into two clinical (P = 0.033) models in the training group, whereas
groups: 464 patients (from 10 centers) were evaluated both radiomics and clinical models showed similar
as the training group and 109 patients (from 1 center) performance (P = 0.613). According to their study, the
were evaluated as the test group. Important variables combined model based on radiomics features and clinical
were determined using the adaptive least absolute and dosimetric parameters can be used to predict 1-year
shrinkage and selection operator (LASSO) method. local tumor control in patients with lung cancer receiving
12
Three prognostic models (clinical, radiomics, and SBRT.
combined) were trained using the random survival forest Isoyama-Shirakawa et al. conducted a study on
(RSF) algorithm. The CI values in the clinical, radiomics, 125 patients with early-stage NSCLC treated with SBRT.
and combined models constructed using clinical and They generated the radiomics score and investigated the
radiomics features were 0.57, 0.55, and 0.61 for LRR and effect of this score in predicting LC and metastasis-free
0.59, 0.67, and 0.68 for DM, respectively. According to survival (MFS). The median BED10 for SBRT was 88.9
their study, although DM could be predicted through the (range: 61.2 – 119.0) Gy. Planning CT-based radiomics
RSF algorithm in patients with early-stage NSCLC who features were generated. From 432 radiomics features,
received SBRT using breath-hold CT-based radiomics five important variables were identified using the LASSO
features, the same algorithm was considered to have a method. The radiomics score was obtained using five
lower potential to predict LRR. 10 significant radiomics features, which was statistically

Volume 1 Issue 4 (2024) 3 doi: 10.36922/aih.3541