Intevo 16

The sentinel node detection was carried out using a 2-day protocol, which involves the injection performed on the afternoon of the day before the surgery and the scintigraphy performed 1 h after the injection. Each patient received 55.5 MBq in 1 ml of Tc-99m phytate delivered peritumorally in each breast. The injection in each breast was given at four sites: two superficial and two deep injections [17 (link)]. One hour after the injection, planar images in 3 directions (anterior, 30° oblique, and 60° oblique views) were taken, and each scan time was 2 min with 512 × 512 matrices [13 (link)]. An additional lateral view was taken when no image was obtained on three views. The SPECT/CT study then followed, with 10 s’ data collection for each step, with 45 steps for 180 degrees of collection (dual cameras) with 128 × 128 matrices, and then, CT was performed to produce axial and coronal fusion images. Both planar imaging and SPECT/CT were performed using a Siemens Symbia Intevo16 (Siemens Healthcare) equipped with a low-energy high-resolution (LEHR) collimator.

Training cohort was based on a population of manually contoured left and right kidneys from varied group of clinical cases. The largest of these was a set of post-treatment ¹⁷⁷Lu-octreotate therapy of neuroendocrine cancer acquired on a hybrid SPECT system with low-dose CT acquisition and 5 mm slice thickness (Siemens Symbia T6 & Intevo 16, Siemens Healthineers, Erlangen, Germany). A subset of patients scanned on dedicated diagnostic CT (Siemens Force, 0.8–5.0 mm slice thickness) and radiotherapy simulation CT systems (Philips Brilliance Big Bore, 3 mm slice thickness, Philips Medical Systems, Cleveland, OH, USA) were included to better adapt the model for detection across different populations and equipment types. A total of 89 manually contoured patients were included for training. Each patient was augmented seven times with a random degree of added noise, edge enhancement, Gaussian smoothing, change in global HU values, translation, and in-plane rotation to avoid CNN overfitting due to non-anatomical image feature (6 ). This provided 712 subjects available for model training. A detailed description of the image augmentation techniques used is given in the Appendix S1 in Supplementary Material.