Purpose: Digitization of interstitial needle is a complicated and tedious process compared with simple applicators in high dose-rate brachytherapy (HDRBT). It becomes particularly challenging when needles are close or touching each other. The accuracy of manual digitization is subject to human experience and time available for this task. To overcome this challenge, we develop a deep-learning assisted auto-digitization tool for interstitial needles, which becomes a new module of our AutoBrachy system.
Methods: Our digitization method consisted two steps. The first step used a neural network with a U-net structure to segment all needles from CT images. We trained the U-net using 95,000 CT slices collected from 15 patient cases and augmented by translation and rotation, and corresponding applicator mask images. The second step used a novel scheme integrating an iterative nearest-neighbor clustering method and a polynomial curve fitting method to separate segmented needle regions into different needles and extract each needleâ€™s central path. Applicator tip position was determined as the point with the highest CT-number gradient along the path. To evaluated out method, we used ten interstitial HDRBT cases that were not used in the U-net training.
Results: For the segmentation step, the average Dice similarity coefficient between automatic and manual segmentation was 0.95 in training, 0.94 in validation, and 0.93 in test cases. After the digitization step, the Hausdorff distance between needles determined by our tool and manually was ~0.6 mm on average and the mean tip position distance was ~0.4 mm. It took about 3 min to complete the digitization of an interstitial HDRBT case.
Conclusion: We have developed a deep-learning assisted auto-digitization tool for interstitial HDRBT. The achieved accuracy and efficiency make our tool clinically attractive.