Automated brachytherapy seed counting with multimodal image registration on trained deep learning convolutional neural network
Session type: Poster / e-Poster / Silent Theatre session
Patient choice and information in (early stage) prostate cancer often includes radiotherapy via iodine-125 seed brachytherapy. Image processing  opens multimodal registration to automatically assist oncology team in postoperative review. Seed count and position is part of RCR information on quality assessment  in dosimetry and informs options to intervene .
Patients undergo CT scan for assessment of implant after one month. Artefactual CT imaging due to seed leads to time-consuming difficult task to review. This work created a CNN deep learning workflow which identified I-125 seeds. Dynocortex is based on a training set derived from (~200) publicly available data supplemented by few (<30) local patients to produce working system advising medical physics experts of seed numbering and location. The system automatically delineates and then co-registers structures in CT, MRI, PET/CT and ULS for quantitative analysis of differences between modalities.
Model count compared to ground truth adequately to flag significant deviation cases for human inspection. Human assessor performs better with training. Current model is limited to <30 cases training and shows useful result at fast (<10second) evaluation.case IDGround truthModel countdifferenceseed15048-2seed26653-13seed36257-5seed47370-3seed968691seed1058591seed118580-5seed126054-6seed139383-10avg6864-5stddev14125
Proof of concept that deep learning was implemented in an oncology team (support from external expertise) exemplifiying NHS guidance and RCR framework  in a timely context. Next steps include discussion with patient representatives and clinical risk assessment.
 Br J Radiol. 2013 Jun; 86(1026): Technical and dosimetric aspects of iodine-125 seed reimplantation in suboptimal prostate implants. L G Marcu, PhD and J M Lawson, MSc