The Ethics Review Committee of Chiba University Hospital approved the present retrospective study. All imaging data used for analysis were necessary for clinical diagnosis, and no examinations were performed for this study. The Ethics Review Committee waived written consent.
Patients
A total of 550 patients (49 males and 501 females; median age, 63 years; range, 33–88 years) on whom bone scintigraphy was performed for clinical diagnosis were included. Of these patients, 500 and 50 were randomly selected for training and obtaining evaluation data, respectively. The breakdown of patient diseases in training and evaluation data is shown in Table 1. The diagnosis of the presence of bone metastases was determined by a radiologist using bone scintigraphy, computed tomography, magnetic resonance imaging, blood tests, and clinical findings.
Table 1
Breakdown of patient diseases in training and evaluation data
Disease | Training data | | Evaluation data |
with bone metastasis | without bone metastasis | Total | | with bone metastasis | without bone metastasis | Total |
Breast cancer | 43 | 383 | 426 | | 10 | 32 | 42 |
Lung cancer | 9 | 40 | 49 | | 3 | 2 | 5 |
Prostate cancer | 10 | 4 | 14 | | 1 | 0 | 1 |
Thyroid cancer | − | 1 | 1 | | 1 | − | 1 |
Rectum cancer | 2 | 2 | 4 | | − | − | − |
Hepatocellular carcinoma | 2 | 2 | 4 | | 1 | − | 1 |
Appendiceal cancer | 1 | − | 1 | | − | − | − |
Plasmacytoma | 1 | − | 1 | | − | − | − |
Total | 68 | 432 | 500 | | 16 | 34 | 50 |
Data acquisition
All patients were examined 3–4 hours after receiving 699–785 MBq of 99mTc-MDP injection solution (PDRadiopharma Inc., Tokyo, Japan). Anterior and posterior whole-body imaging was performed using a gamma camera (NM/CT 870 DR hybrid SPECT/CT scanner; GE Healthcare, Chicago, IL) with a low-energy high-resolution-sensitivity collimator. Whole-body imaging was performed with a matrix size of 1024 x 256, pixel size of 2.21 mm, 15% energy window centered at the photopeak energy (140.5 keV), and bed speed of 13.3 cm/min.
Low-count image preparation
Low-count Original images were created from all Reference images in patient data (100% counts) using the Poisson resampling method [20] application installed in Xeleris 4DR (GE Healthcare, Chicago, IL). The Poisson resampling method arbitrarily subtracts image counts and adds Poisson noise corresponding to the number of counts. Original images with 75%, 50%, 25%, 10%, and 5% counts were created per patient.
Network architecture and training
We have developed a deep learning model based on U-Net [21]. U-Net was developed for image segmentation; however, it also has models that were developed for noise reduction [22–24]. The structure of our deep learning model is shown in Fig. 1. In our model, unsharp masking [25–28] was incorporated before the final output. The unsharp masking sharpens the Output image by adjusting the difference between the Output' image and the smoothed Output' image and adding it to the Output' image. The formula for unsharp masking is shown below.
$${f}_{uij}={d}_{ij}+\delta \sum _{l=-p}^{p}\sum _{m=-p}^{p}{w}_{lm}\left({d}_{ij}-{d}_{i+l, j+m}\right)$$
Where fuij is the pixel value of the Output image at position (i, j), dij is the pixel value of the Output' image at position (i, j), δ is a parameter to adjust the degree of sharpening, and wlm is a Gaussian filter; δ was set to 15. Mean square error (MSE) was used as the loss function for learning. ReLU was the activation function. Adam was the optimizer, and the learning rate was 1.0 × 10− 6. The batch size was 64, and the number of epochs was 1000.
Data analysis
Gaussian-filtered images were obtained by applying a Gaussian filter to the Original image at each count percentage. The size of the Gaussian filter was set to 7 mm, as this setting has been reported to have the best correlation with the Reference image [29]. The present study compared the results of the Original image, Gaussian-filtered image, and Deep learning-filtered (DL-filtered) image, i.e., our model’s output image.
First, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) with the Reference image were calculated for the Original, Gaussian-filtered, and DL-filtered images.
PSNR and SSIM were calculated using the following equations.
\(PSNR=10{\text{log}}_{10}\left(\frac{{I}^{2}}{MSE}\right)\) [db]
Where I is the maximum count of Reference images.
$$SSIM\left(x, y\right)=\frac{\left(2{u}_{x}{u}_{y}+{C}_{1}\right)\left(2{\sigma }_{xy}+{C}_{2}\right)}{\left({\mu }_{x}^{2}+{\mu }_{y}^{2}+{C}_{1}\right)\left({\sigma }_{x}^{2}+{\sigma }_{y}^{2}+{C}_{2}\right)}$$
Where x is the Reference image, y is the target image for comparison, µx and µy are the mean pixel values of x and y, respectively, and σx and σy are the standard deviations of the pixel values of x and y, respectively. σxy is the covariance of the pixel values of x and y. C1 and C2 are constants, expressed as C1=(K1I)2 and C2=(K2I)2, with K1 and K2 set to 0.01 and 0.03, respectively, to avoid division due to minimal values. PSNR and SSIM were compared among Original, Gaussian-filtered, and DL-filtered images.
Furthermore, Reference, Original, Gaussian-filtered, and DL-filtered images were analyzed using BONENAVI software (PDR Pharma, Tokyo). The bone segments analyzed were whole-body bones (skull, cervical spine, thoracic spine, lumbar spine, humerus, thorax, pelvic bone, and femur), except for bones of the peripheral limbs. The artificial neural network (ANN) value, bone scan index (BSI), and hot spot number (Hs) were calculated. The minimum and maximum ANN values were calculated as 0 and 1, respectively; ANN value indicates the confidence level of bone metastasis with a threshold value of 0.5 based on factors such as the shape, location, and count of high accumulated areas [17, 19]. The percentage of bone segment areas with high-risk bone metastasis accumulation sites are shown as BSI and the number as Hs. The analysis values of the Reference image were compared with those of the Original, Gaussian-filtered, and DL-filtered images.
The sensitivity, specificity, and diagnostic accuracy were calculated based on ANN value of > 0.5 indicating the presence of bone metastasis. In addition, receiver operating characteristic (ROC) analysis was performed, and the area under the curve (AUC) was compared.
Statistical analysis
PSNR and SSIM with the Reference image were evaluated using the Steel-Dwass test. ANN values, BSI, and Hs were evaluated using the Steel test, via comparison with those of the Reference image as the control. These values were statistically analyzed using JMP Pro (version 16.1.0; SAS Institute, Cary, NC). DeLong's test was used to evaluate the AUC calculated using ROC analysis. Statistical analyses were performed using EZR (Saitama Medical School Hospital, Saitama, Japan), a GUI of R (The R Foundation for Statistical Computing, Vienna, Austria). Statistical significance was set at p < 0.05 for all analyses.