Deep Learning Based Kidney Segmentation for Glomerular Filtration Rate Measurement Using Quantitative SPECT/CT

Objectives: Glomerular filtration rate (GFR), the rate at which the kidney filters the waste from the blood, is considered the most useful test to measure the level of renal function and determine the stage of kidney disease. Quantitative SPECT/CT is potentially useful for more accurate and reliable GFR measurement than conventional planar scintigraphy [1]. However, manual drawing of a volume of interest (VOI) on renal parenchyma in CT images is labor-intensive and time-consuming job usually taking around 15 min per scan. The aim of this study is to develop fully automated GFR quantification method based on deep learning approach to the three-dimensional (3D) segmentation of kidney parenchyma in CT.

Methods: Two hundred and ninety (290) patients underwent quantitative ^99mTc-DTPA SPECT/CT (GE Discovery NM/CT 670) scans. One-min SPECT data were acquired in a continuous mode 2 min after the intravenous injection of 370 MBq ^99mTc-DTPA. The SPECT images were corrected for attenuation, scatter, and collimator-detector response, and cross-calibrated with a dose calibrator. A nuclear medicine physician drew 2D region of interest (ROI) on renal parenchyma in every 80 to 100 coronal CT slices using vendor’s Q. Metrix software, which provides automatic ROI interpolation between the slices. To reduce the discontinuity in 3D space caused by the 2D ROI drawing, we applied 3D volume smoothing and morphological operations. We used modified 3D U-Net that consists of the contraction and expansion paths and learns an end-to-end mapping between CT and renal parenchyma segmented volumes. Each path has 4 sequential layers composed of a convolution with 3 × 3 × 3 kernels, ReLU (Rectified Linear Unit) and pooling layers (1 × 1 × 1 kernels, Sigmoid for last layer of expansion path). Each layer is updated using the error back-propagation with Adam (adaptive moment estimation) optimizer. Symmetric skip connections between convolutional and up-convolutional layers are used. The U-Net was trained using 240 randomly selected datasets and validated using 50 datasets. Before the training, CT images were down-sampled to 200 × 200 × 160 (2.5 mm³) and cropped into a 144 × 86 × 70 matrix to reduce the input data size. For the quantitative performance evaluation, Dice similarity coefficient between manual drawing and deep learning output was calculated. We calculate % injected dose (%ID) by applying the deep learning output (3D VOI) to the quantitative SPECT images. We also assessed the correlation between the GFR measurements (%ID × 9.1462 + 23.0653) using both segmentation methods [1]. To confirm the consistency of performance, we performed five-fold cross-validation.

Results: We could automatically segment the kidneys in CT images using the proposed method with remarkably high Dice similarity coefficient relative to the manual segmentation (mean ± SD = 0.82 ± 0.054). Although manual segmentation resulted in the discontinuity between slices and vendor’s program sometimes offered wrong ROI interpolation results, the proposed deep learning approach provided 3D kidney parenchyma VOI with no such discontinuity between slices. The GFR values derived using manual and automatic segmentation methods were strongly correlated (R² = 0.94). The absolute difference between the GFR values using manual (45.40±7.48 ml/min) and automatic (46.05±7.32 ml/min) methods was only 3.30±2.89% (left kidney: 3.70±3.16%, right: 3.58±3.18%) in the first cross-validation. The absolute difference obtained in the other cross-validations were 5.75±4.91%, 5.29±3.78%, 3.08±2.81% and 3.41±3.55%, respectively. Conclusion: The proposed deep learning approach to the 3D segmentation of kidney parenchyma in CT enables fast and accurate GFR measurement. Accordingly, this method will be useful for facilitating the GFR measurement using the quantitative SPECT/CT technology.