MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Leveraging Fully-Convolutional Neural Networks to Automatically Delineate Clinical Target Volumes for Spinal Stereotactic Radiosurger

Leveraging Fully-Convolutional Neural Networks to Automatically Delineate Clinical Target Volumes for Spinal Stereotactic Radiosurger

X Wang , C Cardenas , L Court , M Martel , J Yang , C Nguyen , M Amirmazaheri , A Ghia , C Cardenas*, University of Texas MD Anderson Cancer Center, Houston, TX

Presentations

(Sunday, 7/14/2019) 5:00 PM - 6:00 PM

Room: 225BCD

Purpose: Spinal stereotactic radiosurgery (SSRS) is increasingly used to manage spinal metastases. Manual delineations of clinical target volume (CTV) is challenge and time consuming even with consensus guidelines. The goal of this study is to investigate fully-convolutional neural networks to auto-delineate T-spine CTVs.

Methods: Two-hundred and twelve T-spine SSRS cases at our institution were included in this study. Patient’s simulation CT and treatment planning structures were collected; all patients had previously contoured gross tumor volume (GTV) and CTV structures. A two-channel 3D fully-convolutional network was designed to auto-delineate T-spine CTVs. The first channel feeds the network a 3D patch (64x64x64) of the CT image, whereas the second channel (of same size) inputs a patch of the binary GTV mask. The patches are randomly sampled (and centered) about the GTV surface as a function of GTV volume. Our data was split into training (n=106), cross-validation (n=53), and test (n=53) sets. Optimal model parameters were identified by assessing trained model’s predictions on the cross-validation set. Furthermore, post-processing was optimized on the cross-validation set prior to final testing. Overlap and distance metrics are used to quantitatively compare physician and auto-delineations.

Results: Test set’s GTV volumes ranged from 1.1 to 283.3cc (median=18.4cc), whereas physician and auto-delineated CTV volumes ranged from 11.2 to 431.9cc (median=53.6cc) and 3.8 to 491.3 (median=50.2cc), respectively. Average (±std.dev.) Dice Similarity Coefficient, mean surface distance and Hausdorff distance values were 0.728 (±0.121), 3.4mm (2.4mm), and 22.8mm (±15.7mm), respectively. Largest differences between physician and auto-delineated CTVs were observed in the most cranial and caudal extents of the CTV. Predictions for all cases were completed in under 2 minutes.

Conclusion: We developed a T-spine CTV auto-delineation algorithm for SSRS patients. Physician liked its ability in capturing key prospective of our practice and the potential in significantly reduced contouring time even with manual modification.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links