A Multi-Institutional, Convolutional Neural Network-Based Approach to the Detection of Vertebral Body Mis-Alignments in Planar X-Ray Setup Images
R Petragallo1*, P Bertram2, P Halvorsen3, I Iftimia3, J Lamb1, O Morin4, G Narayanasamy5, D Saenz6, K Sukumar7, G Valdes4, L Weinstein8, M Wells7, B Ziemer4, (1) University of California Los Angeles, Los Angeles, CA, (2) Brainlab, Munich, Germany, (3) Beth Israel - Lahey Health, Burlington, MA, (4) University of California San Francisco, San Francisco, CA, (5) University of Arkansas for Medical Sciences, Little Rock, AR, (6) University of Texas HSC SA, San Antonio, TX, (7) Piedmont Hospital, Atlanta, GA, (8) Kaiser Permanente, Oakland, CA
Presentations
MO-FG-BRB-1 (Monday, 7/11/2022) 1:45 PM - 3:45 PM [Eastern Time (GMT-4)]
Ballroom B
Purpose: Mis-alignment to the incorrect vertebral body remains a rare but serious patient safety risk in image-guided radiotherapy (IGRT). Our group has previously presented a single-institution, convolutional neural network (CNN)-based approach for automatic detection of such errors when using a planar stereoscopic x-ray IGRT system. This study expands the previous work to examine the robustness of our model when applied to a larger multi-institutional cohort.
Methods: Digitally reconstructed radiograph (DRR) image pairs were collected from 1,604 treatment fractions of 432 spine radiotherapy patients treated at seven institutions using a stereoscopic x-ray image guidance system. “Off-by-one vertebral body” errors were simulated by translating DRRs along the spinal column using a semi-automated method, and clinically-applied alignments were used for true-negative, “no-error” cases. A leave-one-institution-out approach was used to estimate model accuracy on data from unseen institutions as follows. All of the images from six of the institutions were collected and used to train a CNN model from scratch. The network architecture and hyperparameters were previously tuned using a single institution’s data. Model accuracy was evaluated using images from the seventh institution, which were left out of the training phase entirely. This process was repeated until each institution’s images had been used as the testing dataset.
Results: When the six models were used to classify unseen image pairs from the institution left out during training, the resulting receiver operating characteristic area under the curve values ranged from 0.976 to 0.998. With the specificity fixed at 99%, the corresponding sensitivities ranged from 61.9% to 99.2% (mean: 77.6%). With the specificity fixed at 95%, sensitivities ranged from 85.5% to 99.8% (mean: 92.9%).
Conclusion: This study demonstrated the CNN-based vertebral body mis-alignment model is robust when applied to previously unseen test data from an outside institution, indicating that this proposed additional safeguard against mis-alignment is feasible.
Funding Support, Disclosures, and Conflict of Interest: This work supported in part by AHRQ R01 HS026486. Pascal Bertram is an employee of Brainlab.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email