Prediction of Obliteration After the Gamma Knife Radiosurgery of Arteriovenous Malformations Using Hand-Crafted
David J Wu1, Megan Kollitz2, Mitchell Ward3
1Medicine, University of Minnesota School of Medicine, Minneapolis, USA.
Cureus
|May 24, 2024
Summary
Hand-crafted radiomics and deep learning models can predict outcomes for brain arteriovenous malformation (bAVM) patients undergoing Gamma Knife radiosurgery (GKRS). Both methods showed similar predictive performance, offering promising results for treatment planning.
Area of Science:
- Neurosurgery
- Radiology
- Artificial Intelligence in Medicine
Background:
- Brain arteriovenous malformations (bAVMs) are complex vascular lesions requiring treatment to prevent rupture.
- Gamma Knife radiosurgery (GKRS) is a key treatment modality for bAVMs.
- Predicting treatment outcomes is crucial for patient management.
Purpose of the Study:
- To compare the prediction performance of hand-crafted radiomics and deep learning models for bAVM outcomes after GKRS.
- To identify features that predict favorable versus unfavorable outcomes.
- To evaluate the clinical utility of AI in predicting radiosurgery success.
Main Methods:
- Retrospective review of 42 patients treated with GKRS for bAVMs.
- Development of a Random Forest Classifier (RFC) using hand-crafted radiomic features.
- Fine-tuning a ResNet-34 convolutional neural network (CNN) model.
- Evaluation using ten-fold cross-validation.
Main Results:
- RFC achieved an average accuracy of 68.5% and AUC of 0.705.
- ResNet-34 achieved an average accuracy of 60.0% and AUC of 0.694.
- Four radiomic features significantly discriminated between favorable and unfavorable outcomes.
Conclusions:
- Both radiomics and deep learning models can predict GKRS outcomes for bAVMs with comparable performance.
- Pre-treatment MRI scans are suitable for developing predictive models.
- Further external validation is needed to confirm these promising findings.
