Longitudinal Image Analysis to Predict the Recurrence of Distant Brain Metastases After SRS Treatment
G Szalkowski1, K Thung1, T Zhu2, J Shumway1, J Hall1, C Shen1, P Yap1, J Lian1*, (1) University Of North Carolina At Chapel Hill, (2) Washington University in St Louis, St Louis, MO
Presentations
SU-J-202-4 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]
Room 202
Purpose: To develop a preliminary model to predict if an intracranial recurrence will occur after stereotactic radiosurgery (SRS) treatment based on longitudinal radiomics data of surveillance magnetic resonance imaging (MRI) exams prior to the time of recurrence.
Methods: We retrospectively identified twenty patients with intracranial distant recurrence of brain metastases that had at least two surveillance T1 MR images between their initial SRS treatment and recurrence diagnosis (corresponding to approximately 3 and 6 months before the diagnosis). The screening scans were rigidly aligned to the image used for the treatment of the recurrence. Radiomic features were extracted from areas of untreated brain as well as areas of recurrence (n=37) with a 10mm, 5mm, or no expansion. 92 features were extracted from each of the two scans for 184 features per sample. We evaluated using linear and logistic sparse feature selection to select the most relevant features, and these features were used to train a support-vector machine (SVM) model using either a linear or radial basis unction (RBF) kernel. Model performance was assessed using 10 repetitions of 10 fold cross validation, using the patient identity to perform data splitting.
Results: The combination of logistic sparse feature selection and a linear SVM model gave the highest mean model performance. Using the zero margin data, the model accuracy was 0.95 ± 0.03. Models trained using target contours with a larger margin (allowing for more normal tissue within the contour) still had good predictive power, but were less accurate, at 0.84 ± 0.02 and 0.80 ± 0.03 for the 5mm and 10 mm expansions, respectively.
Conclusion: We trained a model capable of predicting in a given area of the brain based on radiomic features extracted from surveillance MR images taken three and six months before the clinical diagnosis of recurrent intracranial disease.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email