Abstract
267
Objectives: With the recent approvals of new molecular radiotherapy agents, new methods for measurement and assessment of absorbed dose in both normal regions and tumors will be important. Several of these new therapies will require an assessment of dose delivered to the kidneys. In order to do this, the kidneys will have to be accurately delineated, but manual segmentation can be time consuming and burdensome. Automatic methods such as atlas-based segmentation have been developed, but due to variability in patient size, patient shape, and kidney location, it can be challenging to achieve success with these methods. AI-based methods have been shown to produce excellent results when applied to segmentation problems. In this experiment our objective was to show that an AI-based method for segmenting the kidneys would produce clinically acceptable results that could be used for absorbed dose calculations or other applications related to radionuclide therapies.
Methods: Kidney ground truth volumes of interest were contoured manually on 65 anonymized images from various institutions by 6 observers. These images and segmentations were then used to train an in-house neural network. The neural network was based on the well-known U-Net architecture with 3D convolution blocks to better leverage contextual information from all directions. The network takes volumetric data as the input and outputs a probability map for each kidney, which is then binarized. A 5-fold cross validation was performed on the 65 data sets. Dice similarity coefficient (DSC), mean distance to agreement (MDA), and maximum Hausdorff distance (HD) were calculated on the 13-patient test set for each fold.
Results: The segmentation method demonstrated good accuracy with a DSC mean, median, and standard deviation of 0.93, 0.94, and 0.04 respectively. The mean, median, and standard deviation of the MDA was 0.97, 0.81, and 0.50 respectively. The mean, median, and standard deviation of the HD was 10.82, 8.42, and 6.72 respectively.
Conclusions: We found that a 3D convolutional neural network can accurately generate kidney ROIs on CT images. These regions can then be used in radionuclide dosimetry and other applications. This method can be run automatically in the background or during off hours, and therefore has the potential to save time as well as reduce inter-user variability with segmentation tasks. We would like to expand our investigation into these specific areas in the future. We would also like to investigate the utility of this method when applied to other anatomical regions.
