Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Estimation of Tumor Progression for Pediatric Patient with Craniopharyngioma After Surgery and Radiation Therapy Using Machine Learning and Deep Learning Radiomics

W Yang*, C Wang, C Hua, T Davis, J Uh, T Merchant, St. Jude Childrens Research Hospital, Memphis, TN


PO-GePV-M-295 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To estimate the tumor progression rate for pediatric craniopharyngioma patients after surgery, and irradiation therapy using radiomics. The prediction results were compared between handcrafted features analyzed with densely connected artificial neural network (ANN) and automated features analyzed with convolutional neural network (CNN).

Methods: Multiple-center acquired diagnostic MRI with T1-weighted sequence were used in this radiomics study. Eighty-three patients were included based on tumor progression and image availability. Consent forms with Institutional review board approval were obtained. A binary endpoint of tumor progression after treatment was defined as the outcome of the radiomics models. Handcrafted features were input into a densely connected ANN and the MR images with segmentation contour were input into a CNN. Data augmentation was used in both models to balance the positive and negative groups.

Results: The average accuracy of the ANN was 85% with sensitivity and specificity both above 80%. The accuracy of the CNN was 89% with a high sensitivity above 90% while the specificity was slightly below 50%.

Conclusion: Diagnostic MRI based radiomics potentially provide prognostic information on tumor progression with imaging features. Handcrafted features with smaller input dimension outperformed the automated features with CNN based on the relatively small patient cohort. The overfitting of CNN might be mitigated with a larger database.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email