Automatic detection of brachytherapy seeds in 3D ultrasound images using a convolutional neural network.

A novel approach for automatic localization of brachytherapy seeds in 3D transrectal ultrasound (TRUS) images, using machine learning based algorithm, is presented. 3D radiofrequency ultrasound signals were collected from 13 patients using the linear array of the TRUS probe during the brachytherapy procedure in which needles are used for insertion of stranded seeds. Gold standard for the location of seeds on TRUS data were obtained with the guidance of the complete reconstruction of the seed locations from multiple C-arm fluoroscopy views and used in the creation of the training set. We designed and trained a convolutional neural network (CNN) model that worked on 3D cubical sub-regions of the TRUS images, that will be referred to as patches, representing seed, non-seed within a needle track and non-seed elsewhere in the images. The models were trained with these patches to detect the needle track first and then the individual seeds within the needle track. A leave-one-out cross validation approach was used to test the model on the data from eight of the patients, for whom accurate seed locations were available from fluoroscopic imaging. The total inference time was about 7 min for needle track detection in each patient's image and approximately 1 min for seed detection in each needle, leading to a total seed detection time of less than 15 min. Our seed detection algorithm achieved 78% +/- 8% precision, 64% +/- 10% recall and 70% +/- 8% F1_score. The results from our CNN-based method were compared to manual seed localization performed by an expert. The CNN model yielded higher precision (lower false discovery rate) compared to the manual method. The automated approach requires little modification to the current clinical setups and offers the prospect of application in real time intraoperative dosimetric analysis of the implant.