Room: Karl Dean Ballroom C
Purpose: To detect vertebral metastases in simulation CT images for fully-automated palliative radiotherapy treatment planning.
Methods: Simulation CT images from 132 patients were selected to train a deep learning model to predict the presence of metastases in thoracic and lumbar vertebrae. Individual CT slices from palliative radiotherapy plans were labeled as diseased if they contained a physician-contoured GTV or fell within prescription isodose. Slices that fell within prescription isodose were labeled as diseased only if they exhibited visible gross disease. CT slices from breast and lung scans without vertebral disease were labeled as normal. To prepare the input images, the spinal cord was automatically contoured and used as a landmark to crop vertebral bodies. Then, CT slices of adjacent vertebral bodies were placed into three-channel arrays (224x224x3). Using transfer learning, the first 2 layers of the VGG19 deep-learning network were frozen and the remaining layers fine-tuned to return a probability per class (training set: 102 patients, 3741 images; validation set: 39 patients, 2846 images). To assess clinical utility, model prediction for 4 complete patient CTs were compared to a radiation oncologistâ€™s review, which served as ground truth. Accuracy, sensitivity, specificity, and AUC were assessed on a slice-by-slice basis.
Results: The resulting model correctly classified 95% of images in the validation set (sensitivity = 89%, specificity= 96%, AUC=0.98). Model predictions agreed well with the radiation oncologistâ€™s review: accuracy (mean=91%, range=86-96%), sensitivity (mean=95.9%, range=66.7-100%), specificity (mean=90%, range=82-98.2%). High HU values were found on many false positives.
Conclusion: The presence of vertebral metastases can be automatically detected on CT images with high accuracy using a deep learning model. This tool can be used to identify a treatment volume for use in an automated palliative treatment planning system.