Establishment of a deep learning algorithm for dosimetry of radiopharmaceuticals

Abstract

Purpose: The aim of this study was to introduce a volumetric convolutional neural network for segmentation of the kidneys in SPECT images and to apply it in the dosimetry of radiopharmaceuticals of this organ, in order to decrease segmentation time and to standardize the segmentation of the kidneys. Method: Three networks were trained using two network architectures and a total of 216 retrospectively collected images from patients that underwent imaging procedures at Sahlgrenska University Hospital between 2009 and 2018. Hybrid testing and training sets were created containing 55 and 15 image pairs, respectively, from SPECT/CT procedures to be used with the hybrid architecture. Two standard testing sets using diagnostic CT images and CT images from SPECT/CT procedures containing 161 and 53 images, respectively, were compiled to be used with the standard network architecture. Only one testing set was created for this network using the same patients as for the hybrid network, but only using the CT images. Evaluation of the networks was based on the Dice similarity coefficient as well as the activity concentration extracted from the automatically segmented kidneys, in relation to the manually delineated kidneys. Result: Comparison of activity concentration between manually and automatically segmented kidneys presented a percentage difference between -47.66% and 26.66% for all the networks. Kidneys segmented with the hybrid network correlated best with manually segmented kidneys when comparing activity concentration. Conclusion: The hybrid network achieved the most concise results when comparing extracted activity concentration from automatically segmented kidneys with manually segmented ones, while at the same time achieving the lowest Dice coefficient during training. In conclusion, a larger dataset is needed to evaluate the segmentation capabilities in SPECT images.