Page 12 - AIH-1-4

Artificial Intelligence in Health Radiomics in early-stage lung cancer

the training and test sets, respectively. No clinical variable NSCLC at 48 Gy/4 Fr, 50 Gy/4 Fr, or 55 Gy/4 Fr for T1
was predictive of local tumor control in the univariate tumors and 60 Gy/10 Fr or 70 Gy/10 Fr for T2 tumors.
analysis. Similarly, two PET radiomics features and two They extracted 111 radiomics features using PyRadiomics
CT radiomics features did not significantly predict local from both planning CT and PET images. Using three
tumor control. The best predictive models in the training different methods (chi-square test, minimum redundancy
set were obtained by combining one feature from PET with maximum relevance, and ReliefF), they obtained
one feature from CT, achieving 100% sensitivity and 96% 42 important variables and created a model using four
specificity. Another model combining two PET features different algorithms (random forest, SVM, K-nearest
achieved 100% sensitivity and 88% specificity. The second neighborhood, and naive Bayes). SVM with PET radiomics
model achieved an accuracy of 0.91 (sensitivity, 100%; (mean AUC: 0.646), naive Bayes with PET radiomics
specificity, 81%) in the test set. According to their study, (mean AUC: 0.611), and SVM with CT radiomics (mean
two radiomics features derived from 18F-FDG PET were AUC: 0.645) exhibited the highest performance for local
independently associated with local tumor control in recurrence, regional lymph node metastasis, and DM.
patients with NSCLC receiving SBRT and can be combined Their study demonstrated that the model combining PET
into an accurate predictive model. This model can provide imaging features and SVM may be useful for predicting
information regarding local tumor recurrence and assist in local and regional lymph node recurrence and the model
clinical decision-making. 20 combining CT imaging features and SVM may be useful
for predicting distant recurrence. 23
Oikonomou et al. evaluated the recurrence prediction
algorithm using both PET-CT radiomics features and 4. Can radiomics be used in clinical practice?
maximum standardized uptake value (SUVmax) values.
Their study included 150 patients and 172 tumors, and Radiomics can be considered a signature of tumors.
42 features were obtained from CT and PET. There With gradual advancements in technologies in the field
were 11 important variables in the prediction models. of medicine, personalized treatments are becoming
OS, disease-specific survival, and regional control were important. Oncological outcomes are not always similar
estimated in the model established using radiomics in patients at the same stage and age, with the same
features; however, neither SUVmax nor DFS could be performance score, and receiving the same treatment.
predicted using radiomics models. 21 Radiomics is gaining importance in this regard. Tumor
features that are not visible to the clinician can be revealed
In another study, 60% of 464 patients with early-stage through radiomics and play a key role in determining
lung cancer who received SBRT were included in the training the most accurate personalized treatment. Various
set, 40% were included in the test set, and 63 patients from medical imaging technologies, which are non-invasive
another center were included in the external test set. The methods, are used for staging patients with lung cancer
SBRT dose was 40 – 60 Gy administered in 3 – 5 fractions. for treatment selection. Evaluation of medical images
Differences between images from the two centers were is not completely objective and may vary between
eliminated using the ComBat harmonization method. clinicians depending on the person’s experience. Some
A total of 318 radiomics features (106 from each imaging) radiomics features in medical images are not visible to
were obtained from PET, PET-CT, and planning CT using the human eye. As summarized earlier, studies suggest
the PyRadiomics toolbox. In the training and test sets, the that radiomics can be used as a non-invasive adjunct
C-statistics value for predicting regional and/or distant tool for personalized treatment selection and prediction
recurrences using the clinical model was 0.53 – 0.59 (95% of oncological response. Nevertheless, several technical
CI: 0.41 – 0.67), that using the radiomics model was 0.70 difficulties, especially in feature engineering and
– 0.78 (95% CI: 0.63 – 0.88), and that using the combined statistical modeling, and the use of different methods
model was 0.50 – 0.62 (95% CI: 0.37 – 0.69), indicating limit a standardized approach; hence, the clinical use of
that the radiomics model showed the best prediction radiomics still remains under development.
performance. According to this study, radiomics features A standard contouring should be performed for the
obtained from FDG PET-CT before SBRT performed tumor or tumor microenvironment where radiomics
better than clinical parameters in predicting regional and/ features are planned to be studied. Determining the correct
or distant recurrence and determining adjuvant systemic radiomics features may be difficult because of differences
therapy for patients with early-stage NSCLC. 22
between users. Therefore, users must be experienced in
Similarly, Nemoto et al. conducted a recurrence this field and follow accepted guidelines. A well-trained
prediction study using radiomics features from PET AI segmentation system will also help in standardization.
and CT images. They applied SBRT to 82 patients with Furthermore, if the radiation oncologist and radiologist work

Volume 1 Issue 4 (2024) 6 doi: 10.36922/aih.3541