Deep learning-based SPECT/CT quantification of 177Lu uptake in the kidneys

1401 Introduction: The aim of this study was to develop an automated algorithm to quantify activity concentration in the kidneys in Lu-177 SPECT images. The manual routine to quantify activity concentration in the kidneys is accurate but time consuming. Deep learning-based segmentation has been shown to perform well in different modalities but even if the segmentation algorithm performed nearly perfect in the CT the activity concentration would sometimes be measured incorrectly due to misposition of the transferred segmentation VOI (volume of interest) to the SPECT. This misposition between the VOI in the CT and SPECT is due to organ movements. Therefore, an accurate activity quantification with deep learning segmentation had to include adjustment of the created CT VOI to the SPECT. In this study we train a deep convolutional neural network (CNN) to segment the kidneys with respect to the position of the VOI in the CT and SPECT. Methods: The CNN is constructed as a 3D U-net with a dual channel input. The input is the SPECT and CT in a 2x128x128x128 matrix and the output of the network is a 128x128x128 matrix with the segmented kidney VOI and the background. The loss during training is measured with a dice function of the output and the manual segmentation. The CNN was trained on 119 SPECT/CT images and 13 images were used for validation. To validate the accuracy and precision of the method, the activity concentration in the deep learning segmented kidneys in 13 patients were measured and compared to the activity concentration measured in the manually segmented kidneys. Results: The activity concentration measured in the kidneys using the CNN trained on both SPECT and CT differed -2.7% on average compared to the activity concentration measured using the manually segmented kidneys. The largest deviation was -16.5%. The activity concentration measured in the kidneys using the CNN trained on only CT the relative error was 8.2% on average and as most 33%. Conclusions: The SPECT needs to be included to get an accurate and precise automatic segmentation algorithm for SPECT/CT. Regardless of the accuracy and precision of a CT based segmentation algorithm the quantification in SPECT sometimes fails due to organ motion between SPECT acquisition and CT scan.