Reducing scan time in 177Lu planar scintigraphy using convolutional neural network: A Monte Carlo simulation study

Abstract Purpose The aim of this study was to reduce scan time in 177Lu planar scintigraphy through the use of convolutional neural network (CNN) to facilitate personalized dosimetry for 177Lu‐based peptide receptor radionuclide therapy. Methods The CNN model used in this work was based on DenseNet, and the training and testing datasets were generated from Monte Carlo simulation. The CNN input images (IMGinput) consisted of 177Lu planar scintigraphy that contained 10–90% of the total photon counts, while the corresponding full‐count images (IMG100%) were used as the CNN label images. Two‐sample t‐test was conducted to compare the difference in pixel intensities within region of interest between IMG100% and CNN output images (IMGoutput). Results No difference was found in IMGoutput for rods with diameters ranging from 13 to 33 mm in the Derenzo phantom with a target‐to‐background ratio of 20:1, while statistically significant differences were found in IMGoutput for the 10‐mm diameter rods when IMGinput containing 10% to 60% of the total photon counts were denoised. Statistically significant differences were found in IMGoutput for both right and left kidneys in the NCAT phantom when IMGinput containing 10% of the total photon counts were denoised. No statistically significant differences were found in IMGoutput for any other source organs in the NCAT phantom. Conclusion Our results showed that the proposed method can reduce scan time by up to 70% for objects larger than 13 mm, making it a useful tool for personalized dosimetry in 177Lu‐based peptide receptor radionuclide therapy in clinical practice.

Radiation Dose (MIRD) schema, the absorbed dose D to target organ j from source organ k is given by: Gy Bq ⋅ sec Gy Bq ⋅ sec Ã(k) is the cumulated activity in source organ k, which can be derived from 177 Lu planar scintigraphy.7][8][9] By theory, the cumulated activity is obtained by integrating time-activity curve from time 0 to infinity.In our routine clinical practice, time-activity curve is obtained using a hybrid scenario that combines a series of whole body scintigraphies acquired at different time points after the administration of radioactive drug with a single SPECT/CT. 10The SPECT/CT data are used to define volumes-of -interest, which are projected onto whole body planar scintigraphies to measure activity retained in source organs over time.The scan speed for planar scintigraphy is 10 cm/min, so a whole procedure lasts about 20 min for each planar scan.Reducing scan time can increase patient comfort and improve the registration accuracy of whole body planar scintigraphies taken at different time points, which are crucial for precise dose calculation.However, reducing scan time would increase statistical variation and may cause inaccuracy in activity measurements.Deep learning has been used in various imaging applications, and one of them is image denoising. 11,12A more complex model can learn more intricate features from images, but the vanishing gradient problem is a common challenge in training deep neural networks.The dense convolutional network (DenseNet) model proposed by Tong et al. employed dense skip connections to alleviate the vanishing-gradient problem and enhance the feature propagation in deep networks. 13This study investigated the potential of using DenseNet to reduce scan time in 177 Lu planar scintigraphy, in order to facilitate personalized dosimetry for 177 Lu-based peptide receptor radionuclide therapy.Due to the rarity of NETs, the dataset of 177 Lu planar scintigraphy scans from real patients at our institution may not be large enough to build a convolutional neural network (CNN)-based denoising method.Hence, Monte Carlo simulation was used to generate training and testing datasets in this work.

CNN model
The DenseNet model adopted in this study contains one convolution layer to learn low-level features, eight DenseNet blocks to extract high-level features, one bottleneck layer to maintain compactness and reduce computation cost, two deconvolution layers to learn upscaling filters, and one reconstruction layer to generate the output images. 13There are eight convolution layers in each DenseNet block, and all levels of features are combined via skip connections as input for reconstructing the output images.The CNN input images (IMG input ) consisted of 177 Lu planar scintigraphy containing 10% to 90% of the total photon counts, while the corresponding full-count images (IMG 100% ) were used as the CNN label images.The root mean square error (RMSE), also known as the Euclidean distance, was the loss function used in our deep learning model to minimize the difference IMG 100% and IMG input containing 10%−90% of the total photon counts.Using RMSE as the loss function favors a high peak signal-to-noise ratio (PSNR).The DenseNet model was trained from scratch.The filter weights of each layer were initialized by using the MSRA (Microsoft Research Asia) filler. 14All biases were initialized with zero.The models were trained by using Adam (adaptive moment estimation) optimizer with mini-batch size of 32, learning rate of 0.0001, momentum of 0.9, and weight decay of 0.0001. 15

Monte Carlo simulation of 177 Lu planar scintigraphy
A Discovery NM/CT 870 DR (GE Healthcare, Milwaukee, Wisconsin, USA) was modeled by using GATE (GEANT4 Application for Tomographic Emission) 6.0.0. 17The imaging system modeled in this work was a dual-head gamma camera with 3/8 inch NaI(Tl) crystal.The useful field of view (UFOV) of the system was 540 mm (transaxial) × 400 mm (axial). 177Lu planar scintigraphy was simulated by using 20% energy window at 208 keV and medium energy general purpose (MEGP) collimator with hexagonal holes.The MEGP collimator was built with a hole diameter of 3.0 mm, a septal thickness of 1.05 mm, and a hole length of 58 mm.With the MEGP collimator, the system sensitivity was 65 cps/MBq, while the system resolution was 9.4 mm in full width at half maximum (FWHM).A scan time of 5 min was assumed in 177 Lu simulations.Once simulation was completed, the physics list of detected photons was stored in an ASCII (American Standard Code for Information Interchange) file, which recorded time, energy, detector location, and information about the interaction process.The detected photons in list mode were then framed into anterior and posterior projections with pixel size of 2.2 mm × 2.2 mm.

Phantom data for CNN training
Figure 2a shows the Shepp Logan phantom, and Figure 2b shows the Hoffman brain phantom.Both 3D phantoms were first interpolated into a matrix size of 256 × 256 × 150.For each single axial slice, 30 duplicates were created to form an image matrix of 256 × 256 × 30.This process was repeated for all axial slices of both interpolated 3D phantoms, resulting in a total of 300 image sets.These digital phantoms were used to generate simulation data for training the CNN.The 177 Lu planar scintigraphy for these digital phantoms was generated with a uniform water-equivalent attenuation map in GATE Monte Carlo simulation.With regard to activity map, the activity of 150 digital phantoms created based on the Shepp Logan phantom ranged from 2.88 to 20.6 MBq, while the activity of 150 digital phantoms created based on the Hoffman brain phantom ranged from 1.36 to 31 MBq.

Phantom data for CNN testing
The Derenzo phantom shown in Figure 3a had 340 mm diameter and was 78 mm height, so the volume of the Derenzo phantom was approximately 7000 cm 3 .The Derenzo phantom contained six rods with diameter of 10, 13, 17, 22, 28, and 33 mm, and the total volume of these rods was around 180 cm 3 .The 177 Lu planar scintigraphy of the Derenzo phantom was generated with a uniform water-equivalent attenuation map in GATE Monte Carlo simulation.With regard to activity map, the activity concentration of the rods was 10 MBq/kg, so the activity of Derenzo phantom with a target-to-background ratio (TBR) of 5:1 was 15.44 MBq, which was 5.21 MBq for the Derenzo phantom with a TBR of 20:1. Figure 3b shows the NCAT phantom, which provided a realistic model of the human anatomy. 18The 177 Lu planar scintigraphy of the NCAT phantom was generated with tissue compositions recommended by ICRU (International Commission on Radiation Units and Measurements) Report 44 in GATE Monte Carlo simulation. 19With regard to activity map, the time-activity curves expressed in percent injected activity (%IA) for kidney, spleen and liver measured from real patient scans are shown in Figure 3c.

Quantitative analysis
The difference between IMG 100% and CNN output images (IMG output ) was quantified by using RMSE, PSNR, and structural similarity index measure (SSIM).
where V was the number of voxels within the whole image.ROIs were drawn on the rods of Derenzo phantom by using the Otsu thresholding method and were drawn manually on the source organs in NCAT phantom to calculate the mean and SD within ROIs. 20In addition, two-sample t-test was conducted to compare the difference in pixel intensities within ROIs between IMG 100% and the others.A P value less than 0.01 was considered to be statistically significant.The data analysis processes mentioned above were conducted by using MATLAB 7.1 (The Mathworks, Natick, Massachusetts, USA).

CNN training
In Caffe, one training iteration means that one batch has been processed.According to our results, no obvious improvement in RMSE or PSNR was found after 80 000 iterations, so the CNN model trained for 100 000 iterations was used in this work.Figure 4a shows IMG 100% for the Shepp Logan phantom, while Figure 4b shows IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts.The corresponding IMG output and the difference image (IMG diff ) between IMG 100% and IMG output are shown in Figure 4c and 4d, respectively.Figure 5 presents RMSE, PSNR, and SSIM between IMG 100% and IMG output generated by denoising IMG input that contained 10-90% of the total photon counts for the Shepp Logan phantom.Figure 6a shows IMG 100% for the Hoffman brain phantom, while shows IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts.The corresponding IMG output and IMG diff are shown in Figure 6c and 6d, respectively.Figure 7 presents RMSE, PSNR, and SSIM between IMG 100% and IMG output generated by denoising IMG input that contained 10-90% of the total photon counts for the Hoffman brain phantom.Overall, the proposed CNN-based denoising method could reduce the image noise in 177 Lu planar scintigraphy, but IMG output resulted from denoising IMG input that contained fewer photon counts was less similar to IMG 100% .The RMSE, PSNR, and SSIM for the Shepp Logan phantom were similar to those for the Hoffman brain phantom.This finding implied that anatomical complexity did not impact the imaging performance of IMG output seriously.

CNN testing
Figure 8a shows IMG 100% for the Derenzo phantom with 5:1 TBR, while Figure 8b shows IMG input containing 10%, 30%, 50%, 70%, and 90% of the total   mm diameter rods, while no difference was found in any IMG output for 13 mm diameter rods.
The corresponding IMG output are shown in Figure 9c.
Figure 9d-f   changes in noise texture can be observed in IMG 10% output .Moreover, Table 5 shows there were statistically significant differences in IMG 10% output for both right and left kidneys.On the other hand, no statistically significant differences were found in IMG output for any other source organs in the NCAT phantom.However, a P value less than 0.05 was observed in spleen (24 h after injection) and left kidney (24, 96, and 168 h after injection) when comparing IMG 100% and IMG 20% output .Contrarily, all the P values were larger than 0.05 when comparing IMG 100% and IMG 30% output ,indicating that the image denoising procedure should have little impact on the time activity curve derived from 177 Lu planar scintigraphies that contained 30% of the total photon counts.

DISCUSSION
The detection of gamma-emitting radionuclide by planar imaging is a random process that is governed by Poisson statistics.Reducing scan time can increase patient comfort, but it also produces statistical noise in the resulting images and may thus lead to bias in the measurements of organ activity.Image denoising is a task that recovers a clean image from a noisy version.
In nuclear medicine imaging, several physical factors were found to have influence on radioactivity quantification, including spatial resolution, partial volume effect and statistical noise. 21Hence, a successful denoising method for 177 Lu planar scintigraphy should be able to suppress statistical noise without causing serious resolution loss.In general, denoising methods can be divided into two main categories: model-based optimization methods and discriminative learning methods. 22The most actively studied discriminative methods are deep neural networks, in which differences can be observed among various network models.Minarik et al. used a denoising convolutional neural network (DnCNN), which has three convolution layers to reduce statistical noise in whole-body bone scintigraphy and decrease scan time. 23According to their results, the DnCNN enables reducing the scanning time by half and still obtaining good accuracy for bone metastasis assessment.Olia  24 Their results demonstrated that GAN could recover the underlying information in 1/2-dose and 1/4dose SPECT images.Because the architecture of GAN is more complex than that of DnCNN, it may explain why GAN achieves better denoising performance.On the other hand, it is rather challenging to train a GAN model due to vanishing gradient, mode collapse, and nonconvergence.The architecture of DenseNet used in this work is more complex than that of DnCNN.Moreover, the dense skip connections make network training easier and generalization performance better, so this study investigated the potential of using DenseNet to reduce scan time in 177 Lu planar scintigraphy.The DenseNet model proposed by Tong et al. was originally designed for super-resolution, 13 which has been adopted by Kim et al. to enhance signal and noise performance in 99m Tc SPECT imaging. 25Their super-resolution method produced better image quality compared to FBP and OS-EM in terms of contrast-to-noise ratio (CNR), coefficient of variation (COV) and PSNR.Besides resolution recovery, DenseNet has been proven to be an effective type of neural network to reduce the image noise due to low-dose CT scans. 26However, to the best of our knowledge, it has not been used for image denois-ing in 177 Lu planar scintigraphy.Since both image detail and statistical noise are high-frequency components,the process of denoising would inevitably lead to some loss of spatial resolution. 27Therefore, the DenseNet model with dense skip connections was used in this work to evaluate its efficacy in balancing the tradeoff between noise reduction and resolution degradation.
For CNN-based denoising methods, the datasets for model training are crucial to the resulting performance. 23,28Monte Carlo techniques are essential tools in nuclear medicine imaging and have been applied to develop image correction and processing techniques for over 50 years. 29Because simulation can prevent biases introduced into deep learning models due to insufficient training samples or imperfect data quality that are often encountered in real patient scans, simulated 177 Lu planar scintigraphies were used for CNN training and testing in this work.The Shepp Logan phantom and the Hoffman brain phantom are both 3D phantoms but have very different anatomical complexity, so they were used for model training to ensure sufficient data diversity.With regards to model testing, the Derenzo phantom was used to investigate the impact of scan time reduction on the photon counts of hot rods with different sizes, while the NCAT phantom was used to investigate the impact of scan time reduction on the photon counts of source organs measured at for hot rods larger than 13 mm in diameter.Although statistically significant differences were found in IMG 30% output for 10-mm diameter rods, the difference was 4.55% for the Derenzo phantom with 5:1 TBR and 5.51% for the Derenzo phantom with 20:1 TBR.In IMG 50% output , the difference was reduced to 2.57% and 3.60% for the Derenzo phantom with 5:1 TBR and 20:1 TBR, respectively.According to MIRD schema, the activity in source organ k is given by the expression: ).7][8][9] Hence, an accurate determination of the photon counts in Lu-177 planar scintigraphy is crucial for calculating the source activity, which is required for calculating the absorbed dose in target organ j (see Equation 1).Based on NCAT phantom evaluation, the changes in noise texture observed in IMG 10% output were not found in IMG 30% output .Moreover, no significant difference was found in photon counts between IMG 100% and IMG 30% output for source organs measured at all time points.Therefore, the proposed CNN-based denoising method enables a reduction of up to 70% in scan time, or an increase in scan speed by three times.Increasing the scan speed to 30 cm/min in whole-body planar scintigraphy shortens the procedure to approximately 7 min per planar scan.The proposed method has the potential to improve patient comfort while maintaining dosimetry accuracy for 177 Lu-based peptide receptor radionuclide therapy, which may facilitate personalized dosimetry in clinical practice.However, further patient studies are needed to confirm its clinical efficacy.

CONCLUSION
While reducing scan time in nuclear medicine imaging increases patient comfort,it also creates statistical noise in images and affects activity measurement accuracy.

AC K N OW L E D G M E N T S
This research was supported in part by a grant from the Kaohsiung Medical University Research Foundation (KMU-M112023).

C O N F L I C T O F I N T E R E S T S TAT E M E N T
The authors declare no conflicts of interest.

F I G U R E 1
Flowchart of DenseNet training and testing for reducing scan time in 177 Lu planar scintigraphy.IMG x% input : CNN input images that contained x% of the total photon counts; IMG x% output : CNN output images generated by denoising IMG x% input ; IMG 100% : full-count images.

F I G U R E 2 F I G U R E 3
Phantoms for CNN training: (a) the Shepp Logan phantom and (b) the Hoffman brain phantom.Phantoms for CNN testing: (a) the Derenzo phantom, (b) the NCAT phantom, and (c) time-activity curves for the NCAT phantom.

F I G U R E 4
(a) IMG 100% for the Shepp Logan phantom, (b) IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts (from top to bottom) and the corresponding (c) IMG output and (d) IMG diff .F I G U R E 5 (a) RMSE, (b) PSNR, (c) SSIM between IMG 100% and IMG output generated by denoising IMG input , which contained 10% to 90% of the total photon counts for the Shepp Logan phantom.F I G U R E 6 (a) IMG 100% for the Hoffman brain phantom, (b) IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts (from top to bottom) and the corresponding (c) IMG output and (d) IMG diff .

F
photon counts.The corresponding IMG output are shown in Figure 8c. Figure 8d-f shows IMG 100% , IMG input, and IMG output for the Derenzo phantom with 20:1 TBR.The mean and SD of photon counts for the hot rods in Derenzo phantom with 5:1 TBR are summarized in Table1.IMG

F I G U R E 9
(a) IMG100%  for the NCAT phantom with activity distribution measured 24 h after injection, (b) IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts (from top to bottom), and (c) the corresponding IMG output .(d) IMG100% for the NCAT phantom with activity distribution measured 96 h after injection, (e) IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts and (f) the corresponding IMG output .(g) IMG100%  for the NCAT phantom with activity distribution measured 168 h after injection, (h) IMG input containing 10%, 30%, 50%, 70%, and 90% of the total photon counts and (i) the corresponding IMG output .
) I A and I P are the anterior image and the posterior image of the source region in planar scintigraphic imaging ( counts sec ). is the transmission factor, and C is the imaging system calibration factor ( counts∕sec Bq To evaluate the quality and reliability of 177 Lu planar scintigraphy within local sub-regions, region of interest (ROI) analysis was performed to evaluate IMG 100%(low noise), IMG input (high noise), and IMG output (denoising).

Table 1 .
IMG Analysis results of photon counts for the hot rods in the Derenzo phantom with 5:1 TBR.
x% input represented IMG input that contained x% of the total photon counts, while IMG x% output represented IMG output generated by denoising IMG x% input .Moreover, Table 1 also shows the P value obtained by comparing the photon counts of the hot rods in Derenzo phantom with 5:1 TBR between IMG 100% and the others using two-sample t-test.Table 2 summarizes the mean and SD of photon counts for the Derenzo phantom with 20:1 TBR and the P value obtained from two-sample t-test.Based on naked eye observation, the 10 mm diameter rods in Derenzo phantom with 5:1 TBR can be discriminated from one another in IMG TA B L E 1

10 mm 13 mm 17 mm 22 mm 28 mm 33 mm Mean SD P Mean SD P Mean SD P Mean SD P Mean SD P
Analysis results of photon counts for the hot rods in the Derenzo phantom with 20:1 TBR.
TA B L E 2 Analysis results of photon counts for the source organs in the NCAT phantom with activity distribution measured 24 h after injection.
TA B L E 3*The asterisk indicates a P value < 0.01.et al. used a generative adversarial network (GAN) to reduce statistical noise in SPECT myocardial perfusion images, leading to a decrease in patient radiation dose.
Analysis results of photon counts for the source organs in the NCAT phantom with activity distribution measured 96 h after injection.
*The asterisk indicates a P value < 0.01.
Analysis results of photon counts for the source organs in the NCAT phantom with activity distribution measured 168 h after injection., this study investigated the potential of using DenseNet to reduce scan time in177Lu planar scintigraphy.Our results indicate that the proposed CNN-based denoising method can reduce scan time by up to 70% for objects larger than 13 mm.This could greatly facilitate personalized dosimetry in clinical practice for177Lu-based peptide receptor radionuclide therapy.Conceptualization, Ching-Ching Yang, Kuan-Yin Ko, and Pei-Yao Lin; methodology, Ching-Ching Yang, Kuan-Yin Ko, and Pei-Yao Lin; software, Ching-Ching Yang; validation, Ching-Ching Yang; formal analysis, Ching-Ching Yang and Pei-Yao Lin; investigation, Ching-Ching Yang and Kuan-Yin Ko; resources, Ching-Ching Yang and Kuan-Yin Ko; data curation, Ching-Ching Yang, Kuan-Yin Ko, and Pei-Yao Lin; writing, Ching-Ching Yang, visualization, Ching-Ching Yang; supervision, Ching-Ching Yang and Kuan-Yin Ko; project administration, Ching-Ching Yang and Kuan-Yin Ko.
TA B L E 5*The asterisk indicates a P value < 0.01.ThereforeAU T H O R C O N T R I B U T I O N S