Room: AAPM ePoster Library
Purpose: Despite modern image guidance technology, alignment to the wrong vertebral level remains a rare but serious error in radiotherapy delivery. Our group is developing a software failsafe system to mitigate this human error failure mode. Here we compare image classification results from two convolutional neural network (CNN)-based approaches and a logistic regression model for detection of translational shifts of one vertebral body away from the target.
Methods: The dataset consisted of x-ray and digitally reconstructed radiograph (DRR) image pairs from 71 consecutive thoracic spine patients treated at our institution using a stereoscopic onboard image guidance system during the years 2014-2017. To simulate positioning errors, the original DRRs were shifted by one vertebral body in both directions using a semi-automated method. The final dataset was composed of 1,980 x-ray/DRR image pairs. Each image from the stereoscopic system was treated independently. A purpose-built CNN, transfer learning using a pre-trained CNN, and a logistic regression model were all trained using 80% of the dataset, and classification accuracies were evaluated using the remaining 20%.
Results: When the purpose-built CNN was used to classify the previously unseen test image pairs, the resulting receiver operating characteristic area under the curve (AUC) was 0.972. For comparison, the transfer learning and logistic regression models tested on the same dataset obtained AUCs of 0.876 and 0.801, respectively. With the specificity fixed at 99%, the purpose-built CNN achieved a sensitivity of 64.5% in correctly classifying translational shifts of one vertebral body as compared to a sensitivity of 32.3% for the transfer learning model and 23.7% for the logistic regression model.
Conclusion: We have developed a custom deep learning-based algorithm which successfully detects shifts of one vertebral body with a higher degree of accuracy than that achieved by either a transfer learning-based CNN approach or a standard logistic regression model.
Funding Support, Disclosures, and Conflict of Interest: This work supported in part by AHRQ R01 HS026486