Development and clinical implementation of SeedNet: A sliding-window convolutional neural network for radioactive seed identification in MRI-assisted radiosurgery (MARS)

Purpose: To develop and evaluate a sliding-window convolutional neural network (CNN) for radioactive seed identification in MRI of the prostate after permanent implant brachytherapy. Methods: Sixty-eight patients underwent prostate cancer low-dose-rate (LDR) brachytherapy using radioactive seeds stranded with positive contrast MR-signal seed markers and were scanned using a balanced steady-state free precession pulse sequence with and without an endorectal coil (ERC). A sliding-window CNN algorithm (SeedNet) was developed to scan the prostate images using 3D sub-windows and to identify the implanted radioactive seeds. The algorithm was trained on sub-windows extracted from 18 patient images. Seed detection performance was evaluated by computing precision, recall, F₁-score, false discovery rate, and false–negative rate. Seed localization performance was evaluated by computing the RMS error (RMSE) between the manually identified and algorithm-inferred seed locations. SeedNet was implemented into a clinical software package and evaluated on sub-windows extracted from 40 test patients. Results: SeedNet achieved 97.6 ± 2.2% recall and 97.2 ± 1.9% precision for radioactive seed detection and 0.19 ± 0.04 mm RMSE for seed localization in the images acquired with an ERC. Without the ERC, the recall remained high, but the false–positive rate increased; the RMSE of the seed locations increased marginally. The clinical integration of SeedNet slightly increased the run-time, but the overall run-time was still low. Conclusion: SeedNet can be used to perform automated radioactive seed identification in prostate MRI after LDR brachytherapy. Image quality improvement through pulse sequence optimization is expected to improve SeedNet’s performance when imaging without an ERC.