Clinical performance of deep learning enhanced ultra-fast whole-body scintigraphy in patients with suspected malignancy

doi:10.21203/rs.3.rs-4721424/v1

Download PDF

Research Article

Clinical performance of deep learning enhanced ultra-fast whole-body scintigraphy in patients with suspected malignancy

https://doi.org/10.21203/rs.3.rs-4721424/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

To evaluate the clinical performance of two deep learning methods, utilizing real clinical pairs and simulated datasets, for fast whole-body scintigraphy.

Methods

This prospective study enrolled 83 patients with suspected bone metastasis. All patients received SPECT whole-body scintigraphy (WBS) at the speed of 20cm/min (1x), 40cm/min (2x), 60cm/min (3x). Two deep learning models were introduced to generate high-quality images from fast scans, designated as 2x-real, 3x-real (from real model), and 2x-simu, 3x-simu (from simulated model). A 5-point Likert scale was utilized to evaluate the image quality of each acquisition. Accuracy, sensitivity, specificity, and ROC-AUC were used to evaluate the diagnostic efficacy. Learned perceptual image patch similarity (LPIPS) and fréchet inception distance (FID) were used to assess image quality. Additionally, count-level consistency of WBS was also compared.

Results

Subjective assessments indicated that 1x images exhibited the highest general image quality (Likert score: 4.40 ± 0.45). 2x-real, 2x-simu and 3x-real, 3x-simu images displayed significantly superior quality than those of 2x and 3x image respectively (Likert scores: 3.46 ± 0.47, 3.79 ± 0.55 vs. 2.92 ± 0.41, P < 0.0001;. 2.69 ± 0.40, 2.61 ± 0.41 vs. 1.36 ± 0.51, P < 0.0001). Notably, the quality of 2x-real images was inferior to those of 2x-simu (Likert scores: 3.46 ± 0.47 vs. 3.79 ± 0.55, P = 0.001). The diagnostic efficacy of 2x-real, 2x-simu was indistinguishable from 1x image (accuracy: 81.2%, 80.7% vs. 84.3%; sensitivity: 77.27%, 77.27% vs. 87.18%;specificity: 87.18%, 84.63% vs. 87.18%. all P > 0.05), while 3x-real, 3x-simu had better diagnostic efficacy than 3x (accuracy:65.1%, 66.35% vs. 84.3%; sensitivity: 63.64%, 63.64% vs. 87.18%; specificity: 66.67%, 69.23% vs. 87.18%, all P < 0.05).. Objectively, both real model and simulated model significantly enhanced image quality from the accelerated scans (FID: 0.15 ± 0.18, 0.18 ± 0.18 vs. 0.47 ± 0.34, P < 0.05༛LPIPS: 0.17 ± 0.05, 0.16 ± 0.04 vs. 0.19 ± 0.05, P < 0.05). The count-level consistency with the 1x images was excellent for 2x-real, 3x-real, 2x-simu, and 3x-simu (P < 0.0001).

Conclusions

The ultra-fast 2x speed (real and simulated) image could achieve comparable diagnostic value to those of standard acquisition, and the simulation algorithm could not necessarily reflect the real data.

SPECT

whole-body scintigraphy

deep learning

diagnostic efficacy

Whole body scintigraphy (WBS) is an invaluable tool for the examination of pathological skeletal conditions, such as infectious, traumatic, neoplastic, or other etiologies, by using phosphates labeled with Tc that are rapidly absorbed and cleared in the skeleton[1, 2]. WBS can detect significant changes in bone metabolism weeks to months before evident changes appear on conventional radiographic images, thus demonstrating high sensitivity[3]. Given its high sensitivity, affordability, and the ability to provide an overview of the entire skeletal system in a single imaging session, WBS is the most common examination in nuclear medicine[4]. In 2019, the number of WBS examinations exceeded more than 60% of the total single photon emission computed tomography(SPECT) scans in China[5].

The recommended scanning speed is 10–15 cm/min for WBS to ensure that both anterior and posterior WBS contain more than 1.5 million counts, necessary for achieving diagnostic-quality image resolution[1, 2, 6]. This indicates that a standard WBS typically takes 15–20 minutes, posing a challenge for patients with bone pain or those unable to maintain the same position for an extended period. Therefore, there is an urgent need to accelerate scanning speed to increase examination throughput. Decreasing radioactive tracer dosage is meaningful for patients requiring follow-up WBS with suspected or confirmed bone metastatic. However, lowering scan time and radioactive tracer dosage would result in a decreased number of acquired counts, leading to reduced image quality, increased data noise, potential artifacts in reconstructed image, and deterioration of diagnostic performance[7].

Over recent years, deep learning methods have achieved significant success in various medical imaging processing tasks, including automatic lesion detection and anatomical localization[8], structure segmentation[9], automatic diagnosis[10–12], disease classification and image enhancement[13]. Predominant deep learning methodologies employ a supervised learning approach, where paired input and output data are necessary for effective training of models. Specifically, in the context of accelerating WBS, the procurement of perfectly matched pairs of fast scan images and their corresponding standard scans present a significant challenge. Existing researches can be broadly classified into two approaches: one that relies on the utilization of simulated datasets for model training and another that addresses the inherent mismatch between two distinct real-world acquisitions. The simulation approach encompasses techniques such as generating low-count images from their high-count counterparts or creating simulated standard scans from rapid acquisition processes[14–16]. Conversely, our previous work confronts the challenges posed by the inherent discrepancies observed between two distinct real-world data acquisitions. A RRDB based network and data pre-processing procedures like bladder exclusion and patch-focused registration were introduced to generate high-quality standard scans[17].

In this prospective work, we would like to verify the feasibility of deep learning based fast WBS from a clinic diagnostic view and compare the performance of two models trained using simulation datasets and real clinical pairs.

2.1 Patients

The study was approved by the Institutional Review Board of Shanghai East Hospital (EC. D(BG). 009. 02.1). All patients have signed the informed consent forms. This prospective study enrolled patients with known or suspected bone metastatic lesion, who underwent clinical technetium 99m-methylenediphosphonate (Tc-99m MDP) SPECT WBS at Shanghai East Hospital from August 3 to December 15, 2023.

2.2 Image acquisition

All patients were injected with 9–11 MBq kg − 1 (0.24–0.29mCi kg − 1) Tc-99m MDP, and WBS was performed approximately 3–4 h post-injection. The imaging was performed using the SPECT/CT (Siemens Symbia Intevo, Erlangen, Germany). The SPECT/CT was equipped with a low-energy, high-resolution parallel-hole collimator. The matrix size was 256×1024. The energy peak was centered at 140 keV with a 15% window. WBS images were obtained at three scanning speeds: 20cm/min for standard scans (1x), 40cm/min for 2x fast scans, and 60cm/min for 3x fast scans, with scanning duration of 10min, 5min, and 3.3min respectively. The 3x fast scan was conducted first, followed by the 2x fast scan and the standard scan at last. The scans were consecutive and completed within 20 minutes.

2.3 Deep learning workflow

Two supervised deep learning models were developed for the research. One model was based on the real clinic acquisition pairs (real model). Data pre-process including bladder exclusion, count-level consistency and patch-focused registration were followed the instruction of previous work[17]. Another model was based on the simulation dataset (simulation model). Full count data from the training dataset for real model were also used as the labels. Low count input images were simulated using Poisson resample technique as described in simulation work[14].

Two models shared the same RRDBNet based network architecture, which consist of 9 adjacent dense forward connection modules shown in Figure. 1. An Up-sampling Efficient Sub-Pixel Convolutional Neural Network (ESPCN) and a down-sampling convolutional layer were followed to obtained the refined feature map.

The loss function consisted of four components, including pixel-wised L1 loss, binary cross entropy loss, perceptual loss and generative adversarial network (GAN) loss.

$$\:\begin{array}{c}L={L}_{p}+\lambda\:{L}_{1}+{\eta\:L}_{BCE}+\gamma\:{L}_{d}\#（1）\end{array}$$

Where $\:{L}_{p}$ stands for perceptual loss, $\:{L}_{1}$ is L1 loss, $\:{L}_{BCE}$ is BCE loss and L_d is GAN loss. $\:\lambda\:$, $\:\eta\:$ and $\:\gamma\:$ are constant coefficients that adjust the impact weight of respective loss. In the training process, $\:\lambda\:$, $\:\eta\:$ and $\:\gamma\:$ are set to 0.005, 0.01 and 0.001 respectively.

The networks were trained using Adam optimizer for 800 epochs with a consisting learning rate of 2e-4. Beta1 and Beta2 for Adam optimizer were set to 0.9 and 0.999. Batch size was set to 1 and patch size was set to 192.

2.4 Image analysis

2.4.1 Subjective assessment

In the assessment of WBS, a subjective approach was employed to evaluate specific aspects related to image quality and diagnostic reliability. A 5-point Likert scale was utilized for the assessment by two radiologists with 5 and 10 years' experience. The overall clarity and quality of the images were assessed to ensure that the WBS adequately represented bone structures and abnormalities. The scoring criteria were as follows: 5 points: Excellent image quality with minimal noise and superior display of anatomical details; 4 points: Good image quality with low noise, clear visualization of anatomical structures, with all structures distinctly discernible without impacting the observation and diagnosis of lesions; 3 points: Moderate image quality with medium noise, average display of anatomical structures, minor impact on lesion observation but sufficient for clinical diagnosis; 2 points: Subpar image quality with significant noise, incomplete visualization of anatomical structures, poor display of anatomical details, affecting the observation of lesions; 1 point: Poor image quality with severe noise, incapable of providing useful information, unsuitable for diagnostic purposes. Benign and malignant lesions were determined based on pathological diagnosis, imaging examinations(standard WBS, CT, MRI, PET/CT), and clinical follow-up data (For example, the lesion became larger during image follow-up and the tracer of the lesion was more concentrated than before was considered as malignant). The two radiologists independently read randomly and blindly presented 1x, 2x, 3x, 2x-real, 3x-real, 2x-semi, 3x-semi, and determined whether the lesions were benign (negative) or malignant (positive), while an image quality score of 1 was considered negative. If the results were inconsistent, another senior physician was included to make a judgment. The results of the models were evaluated by followed parameters, including accuracy, sensitivity, specificity, and ROC-AUC.

2.4.2 Objective assessment

For objective assessment, we assessed image quality using LPIPS and FID. Additionally, count-level consistency of WBS was also compared. As described in[17], traditional metrics like peak signal-to-noise ratio (PSNR) and structure similarity index measurement (SSIM) could not fully represent the consistency of WBS images. In this study, FID and LPIPS were used to evaluate the performance of different models: FID is defined as

$$\:\begin{array}{c}FID={\left|\left|\mu\:-{\mu\:}_{w}\right|\right|}^{2}+tr\left(\varSigma\:+{\varSigma\:}_{w}-{2\left(\varSigma\:{\varSigma\:}_{w}\right)}^{\frac{1}{2}}\right)\#\left(2\right)\end{array}$$

where $\:{\mu\:}_{w},\:\:{{\Sigma\:}}_{w}\:$and $\:{\mu\:},\:\:{\Sigma\:}$are the mean and covariance of the real data and model distributions, respectively. Lower FID means smaller distances between synthetic and real data distributions.

LPIPS essentially computes the similarity between the activations of fast scan and standard scan patches for a pre-defined network at multiple level of feature maps. This measure has been shown to match human perception well. A low LPIPS score means that image patches have a closer feature distance and are more perceptual similar. TorchMetrics is used to calculate the score with pretrained vgg19 model. Additionally, count-level consistency of the largest lession in WBS was also compared.

2.5 Statistical analysis

Statistical analysis were performed using GraphPad Prism (8.0.0) and SPSS Statistics version 26 (IBM, Armonk, NY, USA). The quantitative parameters (LPIPS, FID) and subjective scores of image quality were compared with the Wilcoxon signed-rank test. The McNemar test was used to assess the differences in diagnostic performance. Bland–Altman and Pearson correlation analysis were used to demonstrate the consistency of the counts between the 2x-real, 3x-real, 2x-simu, 3x-simu, and 1x images. Statistical significance was considered at a p-value of less than 0.05.

In total, 83 patients (female/male:43/40; age range:25–94 years; BMI range:15.9–26.1) were enrolled in this study. 39 patients were diagnosed with bone metastasis eventually. Their demographic and clinical data are detailed in Table 1.

Table 1

Patient clinical characteristics
N = 83	Overall
Age
Mean(SD)	61 (13)
(Min, Max)	(25, 81)
Gender
Male	40
Female	43
BMI
Mean (SD)	22.5 (1.2)
(Min, Max; 95%CI)	(15.9, 26.1; 21.8–24.1)
Diagnosis
Lung cancer	22
Breast cancer	16
Prostate cancer	15
Gastric cancer	9
Colon cancer	8
Liver cancer	1
Nasopharyngeal carcinoma	5
Bone pain of unknow origin	7

As illustrated in Figure. 2, the 1x images showed best general image quality (Likert score: 4.40 ± 0.45), while 3x signified poor quality (Likert score: 1.36 ± 0.51). 2x-real, 2x-simu and 3x-real, 3x-simu images showed significantly superior to 2x and 3x image in subjective image quality (Likert score: 3.46 ± 0.47, 3.79 ± 0.55 vs. 2.92 ± 0.41; 2.69 ± 0.40, 2.61 ± 0.41 vs. 1.36 ± 0.51. P<0.0001), respectively. The 2x-real image quality is lower than the 2x-simu (Likert score: 3.46 ± 0.47 vs. 3.79 ± 0.55. P = 0.001), and there was no significant difference in 3x-real and 3x-simu image quality (Likert score: 2.69 ± 0.40 vs. 2.61 ± 0.41. P = 0.95).

Figure. 3 and Figure. 4 displayed the anterior WBS images of representative patients with bone metastasis at different acquisition times (1x, 2x, 3x) and deep learning model introduced images (2x-real, 3x-real, and 2x-simu, 3x-simu). A (1x) had the best image quality with minimal noise. The 2x-real(D), 2x-simu(F), 3x-real(E), and 3x-simu(G) images accelerated using deep learning techniques exhibited superior quality compared to the original 2x(B) and 3x(C) images.

Table 2 showed the accuracy, sensitivity, and specificity of 1x, 2x, 2x-real, 2x-simu, 3x, 3x-real, and 3x-simu images. From Table 2, it could be seen that 1x had the highest accuracy, sensitivity, and specificity, followed by 2x-real and 2x-simu, which were improved compared to 2x. The diagnostic efficacy parameters of 3x-real and 3x-simu were lower than 2x but higher than 3x. Table 3 showed that there was no significant statistical difference in the accuracy, sensitivity, and specificity of 2x, 2x-real, and 2x-simu compared to 1x (P > 0.05), and there was no significant statistical difference in the three diagnostic parameters of 2x-real and 2x-simu(P > 0.05). The accuracy, specificity, and sensitivity of 3x were significantly lower than those of 1x (P < 0.05), while the accuracy and specificity of 3x- real were significantly lower than those of 1x (P < 0.05). The accuracy of 3x-simu was significantly lower than that of 1x (P < 0.05). There was no significant difference in the diagnostic efficacy parameters of the three groups between 3x-real and 3x-simu (P > 0.05). Figure. 5 showed the ROC curves of 7 groups. It could be seen that the AUC increased according to 3x, 3x-real, 3x- simu, 2x, 2x-simu, 2x-real, and 1x, and the bases under the curves of 1x, 2x- real, and 2x-simu were all above 0.80.

Table 2

Clinical diagnostic efficacy of 1x, 2x, 3x, 2x-real, 2x-simu, 3x-real, 3x-simu image (N = 83)
Group	Accuracy	Sensitivity	Specificity
1x	84.3% (70/83)	87.18% (34/39)	81.82% (36/44)
2x	73.4% (61/83)	78.38% (29/37)	69.57% (32/46)
2x-real	81.2% (68/83)	89.47% (34/38)	75.56% (34/45)
2x-simu	80.7% (67/83)	85.00% (34/40)	76.74% (33/43)
3x	59.0% (49/83)	64.71% (22/34)	55.10% (27/49)
3x-real	65.1% (54/83)	70.27% (26/37)	60.87% (28/46)
3x-simu	66.3% (55/83)	69.23% (27/39)	63.64% (28/44)

Table 3

Statistics analysis about the accuracy, sensitivity, and specificity of 1x, 2x, 3x, 2x-real, 2x-simu, 3x-real, 3x-simu image (N = 83)
Group	Accuracy	Sensitivity	Specificity
2x vs. 1x	P = 0.087	P = 0.308	P = 0.176
2x-real vs. 1x	P = 0.678	P = 0.754	P = 0.471
2x-simu vs. 1x	P = 0.361	P = 0.780	P = 0.559
2x-real vs. 2x-simu	P = 0.842	P = 0.555	P = 0.896
3x vs. 1x	P = 0.004	P = 0.023	P = 0.006
3x-real vs. 1x	P = 0.004	P = 0.071	P = 0.029
3x-simu vs. 1x	P = 0.007	P = 0.055	P = 0.056
3x-real vs. 3x-simu	P = 0.870	P = 0.921	P = 0.787

Objective assessment of models' performance was shown in Table 4. Both real model and simu model showed significant improvement from the fast scan (P < 0.0001). No difference was found between 2x-real, 2x-simu and 3x-real, x3-simu in FID (P = 0.4599). But 3x-real achieved a better LPIPS performance compared to that of 3x-simu (P = 0.0016).

Table 4

Objective quantitative evaluation
	FID		LPIPS
	2x	3x	2x	3x
Fast Scan	0.465 ± 0.337	0.975 ± 0.585	0.190 ± 0.046	0.231 ± 0.040
Real	0.152 ± 0.183*	0.193 ± 0.227*	0.166 ± 0.046*	0.176 ± 0.046*
Simu	0.177 ± 0.184*	0.200 ± 0.224*	0.163 ± 0.043*	0.186 ± 0.045*†
*: significant difference compared to the fast scan. †: significant difference compared to the Real

Bland–Altman analysis showed a max count value difference of − 0.0055, and the 95% limits of agreement (-19.26, 34.70) contained 37/39 of the difference in the 1x versus 2x-real image (Figure. 6A), (-26.22, 24.68) contained 38/39 of the difference in the 1x versus 2x-simu image (Figure. 6B), (-36.92, 52.92) contained 38/39 of the difference (Figure. 6C) in the 1x versus 3x-real image, (-40.25, 31.94) contained 36/39 of the difference (Figure. 6D) in the 1x versus 3x-simu image. Pearson correlation analysis unveiled a noteworthy association between the counts of 2x-real and 1x (Figure. 7A), exhibiting a sparse correlation coefficient of r = 0.999 (Figure. 7B), a significant correlation between the counts of 2x-simu and 1x, with r = 1.017 (Figure. 7C), a significant correlation between the counts of 3x-real and 1x, with r = 0.953. Additionally, the counts of 3x-simu and 1x exhibit a significant correlation with r = 1.042 (Figure. 7D). (P<0.0001).

The current study validates the efficacy of deep learning methods in enhancing the image quality of 2x images, as demonstrated by both subjective and objective assessments, meeting clinical diagnostic standards. However, Likert scores of 2x-real and 2x-simu exhibit significant differences compared to standard 1x images, which often reflecting subjectivity and reliance on physician experience for interpretation. Nevertheless, objective evaluations using image quality metrics like LPIPS and FID show no significant differences from 1x images. FID quantifies the disparities between synthetic and authentic data distributions, while LPIPS gauges feature distance and perceptual likeness of image patches, potentially indicating consistency in whole-body bone images. Moreover, the observed differences in 2x-real and 2x-simu images do not impact diagnostic efficacy, as their diagnostic accuracy (81.2%, 80.7%), sensitivity (89.47%, 85.00%), specificity (75.56%, 76.74%), and AUC-ROC values (0.82, 0.81) remain comparable to those of 1x images. The results are also similar to those reported in a previous meta-analysis[3], which showed a sensitivity of 86.5% and specificity of 79.9% for diagnosing bone metastases using WBS, affirming their capability to discern malignant bone lesions effectively.

It is noteworthy that in this study, the diagnostic performance of 2x images, with doubled scanning speed, exhibited a slight decrease compared to standard 1x images. But this difference was not statistically significant, possibly due to the relatively small sample size and potential subjective bias in physician interpretation, given the reliance of bone scintigraphy interpretation on individual knowledge and experience. Nonetheless, the AUC-ROC value was below 0.8, indicating a diminished ability to differentiate malignant bone lesions compared to 1x images and 2x-real, 2x-simu images[18]. The image quality of 3x-real and 3x-simu images enhanced by deep learning methods was notably superior to that of 3x images according to both subjective and objective assessments. However, the accuracy, sensitivity, and specificity were significantly lower compared to 1x images, with all AUC-ROC values below 0.7, indicating an insufficient capability to meet clinical diagnostic requirements.

A previous study developed a deep-learning noise reduction (DLNR) algorithm for whole-body cadmium and zinc telluride (CZT) SPECT images using data from 19 patients, demonstrating that image quality remained at a good-to-excellent level even when reducing the acquisition time to 60%[19]. In this current study, a dual-head gamma camera is used to develop deep-learning models for WBS, which is more commonly used in clinical practice[20]. This study boasts a relatively large sample size and shows that ultra-fast imaging at 50% of the acquisition time (using both real and simulated datasets) can achieve comparable diagnostic efficacy to standard acquisition, as determined by subjective and objective assessments. Similarly, Minarik et al. conducted a visual evaluation of bone metastases presence in DL-filtered images with 50% counts and reported no significant difference in diagnostic performance compared to the reference image[21]. A previous study indicated that the evaluated noise-reducing Pixon algorithm enhanced planar processing does not fully compensate for the loss of counts associated with reducing the scan-time in half for WBS, and the algorithm is based on the principle that the ideal image is represented by the lowest possible number of parameters that correctly represent the raw data image[22]. In this study, two supervised deep learning models were trained based on resampled simulation dataset or acquired real clinical pairs. Our results suggested that advancements in deep learning technology have further improved the quality of low-dose scan images.

Our previous study demonstrated that ultrafast SPECT/CT with a 1/7 acquisition time could be improved using the deep learning methods to achieve image quality and diagnostic value comparable to those of standard acquisition protocols[23, 24]. While WBS and bone SPECT imaging both rely on single photon counting principles, adjacent layers and corresponding CT in tomography provide more priori image information, thus, images obtained with 1/7 standard acquisition time can be accelerated using deep learning techniques to meet diagnostic requirements. For instance, gathering additional stratified standard scan times at intervals like 48%, 46%, and 44% could help refine the minimum scan time for accelerated imaging that ensures both image quality and meets clinical requirements. However, in clinical practice, it takes effort to acquire more low-count images. A study generated low-count original images (75%, 50%, 25%, 10%, and 5% counts) from reference images (100% counts) using Poisson resampling indicated that deep learning method improved image quality and bone metastasis detection accuracy for low-count bone scintigraphy[14]. In our study, subjective scores between 2x-simu (low-count original images from reference images Poisson resampling) and 2x-real still exhibited differences, but did not differ significantly in objective LPIPS and FID scores. While objective LPIPS scores between 3x-simu (low-count original images from reference images Poisson resampling) and 3x-real showed significant difference. It is worth considering whether the objective image quality differences between simulated and real images increase with simulated count reductions, emphasizing the need for further investigation with larger sample sizes to validate these observations. As showed in Figure. 3, the image quality of 2x-real (D), 2x-simu (F), 3x-real(E), and 3x-simu (G) is improved compared to 2x (B) and 3x (C) with reduced noise and increased radiopharmaceutical counts. In WBS, there appears to be no significant difference in image quality between 2x-real (D) and 2x-simu (F), as well as between 3x-real (E) and 3x-simu (G). However, in the enlarged images within the red dashed box, it can be observed that the range of bone metastatic lesions in 2x-simu (F) and 3x-simu(G) is closer to the standard 1x (A) image, whereas in 2x-real and (D) 3x-real (E), the lesion range is slightly larger with less distinct boundaries between lesions and normal bone. These findings suggest that simulation images may not accurately simulate real images. While two models shared the same network architecture and training parameters, the difference in the generated images came from the discrepancy of simulated low-count data and real scanned low-count data. Possible explanations are continuous drug metabolized during the examination and the noise not fully consistent with the Poisson down sampling process for atomic decay.

The advancement of SPECT/CT technology has facilitated quantitative evaluations in bone imaging[25–27]. Several studies have shown strong inter-observer agreement in quantitative analysis and established a significant correlation between standardized uptake values (SUV) derived from SPECT images of bone metastases and those from PET images[28–30]. The SUV in SPECT/CT imaging is based on single photon counting, therefore, in WBS, single photon counts also holds certain quantitative reference value. One of the deep learning principles utilized in this study is to improve image quality by increasing the radiopharmaceutical counts in images acquired with reduced scan times. Consequently, the accelerated counts of 39 lesions in 2x-real, 2x-simu, 3x-real, and 3x-simu images exhibit high consistency with the original images. This further confirms the applicability of deep learning-accelerated planar bone imaging in clinical practice from an objective perspective. These findings align with those of[14], where lesion counts in their results are strongly correlated with those of the original standard images, albeit with a 10% decrease in counts, possibly attributed to the risk of signal loss due to smoothing observed in that study. The highly consistency of max lesion count between deep learning enhanced images and reference WBS could benefit from the count consistency preprocess. Hence, the scale gap between low-count scans and full-count scans in different case was well normalized. Our proposed method also contributes to well alignment of max lesion count.

This study has several limitations. Firstly, the prospective nature of data collection from a single center and a relatively small sample size may lead to biased results. Secondly, this study exclusively utilized 99mTc-MDP. It is recommended to explore the use of other radiopharmaceuticals in future studies to enhance the generalizability and robustness of this deep learning models. Lastly, our study just collected images acquired at 2x and 3x scanning times and assessed the quality of these images after undergoing deep learning processing. The results indicated that 2x-real and 2x-simu images meet clinical diagnostic requirements. However, further refinement of the scanning time between 2x and 3x is needed to identify the shortest scanning time that can meet clinical diagnostic needs after using deep learning methods.

Our study indicated that the ultra-fast 2x speed (both real and simulated) images could achieve comparable diagnostic value to those of standard acquisition in subjective and objective assessments, and the simulation algorithm could not necessarily reflect the real data.

SPECT: Single-photon emission computed tomography

WBS: Whole-body scintigraphy

LPIPS: Learned perceptual image patch similarity

FID: Fréchet inception distance

PSNR: Peak signal-to-noise ratio

SSIM: Structure similarity index measurement

SUV: Standardized uptake values

Human Ethics and Consent to Participate declarations: The study was approved by the Institutional Review Board of Shanghai East Hospital (EC. D(BG). 009. 02.1). All patients have signed the informed consent forms.

Consent for publication: Not applicable.

Availability of data and material: The datasets generated or analyzed during this study are available from the corresponding author on reasonable request.

Competing Interests: The authors report no conflicts of interest.

Funding: The study was partially supported by Key Discipline Construction Project of Shanghai Pudong New Area Health Commission (Grant No.PWZxk2022-12); Youth Science and Technology Project of Shanghai Pudong New Area Health Commission (Grant No.PW2023-B08).

Authors’ Contributions: All authors were involved in the study conception and design. Na Qi, Qingyuan Meng, Yihong Yang, Jie Ding and Jun Zhao were involved in the acquisition of data. Na Qi, Boyang Pan, Zengbei Yuan and Nan-jie Gong were involved in analysis of the data. Na Qi was involved in the drafting of the manuscript. All authors were involved with critical revisions of the manuscript.

Acknowledgements: Not applicable.

Van den Wyngaert T, Strobel K, Kampen WU, Kuwert T, van der Bruggen W, et al. The EANM practice guidelines for bone scintigraphy. Eur J Nucl Med Mol Imaging 2016, 43(9):1723-1738.
Bombardieri E, Aktolun C, Baum RP, Bishof-Delaloye A, Buscombe J, et al. Bone scintigraphy: procedure guidelines for tumour imaging. Eur J Nucl Med Mol Imaging 2003, 30(12):BP99-106.
Yang HL, Liu T, Wang XM, Xu Y, Deng SM. Diagnosis of bone metastases: a meta-analysis comparing (1)(8)FDG PET, CT, MRI and bone scintigraphy. Eur Radiol 2011, 21(12):2604-2617.
Kosuda S, Kaji T, Yokoyama H, Yokokawa T, Katayama M, et al. Does bone SPECT actually have lower sensitivity for detecting vertebral metastasis than MRI? J Nucl Med 1996, 37(6):975-978.
Wang J LS. A brief report on the results of the national survey of nuclear medicine in 2020. Chin J Nucl Med Mol Imaging 2020, 49:747-749.
Bartel TB, Kuruva M, Gnanasegaran G, Beheshti M, Cohen EJ, et al. SNMMI Procedure Standard for Bone Scintigraphy 4.0. J Nucl Med Technol 2018, 46(4):398-404.
Schaefferkoetter J, Yan J, Moon S, Chan R, Ortega C, et al. Deep learning for whole-body medical image generation. Eur J Nucl Med Mol Imaging 2021, 48(12):3817-3826.
Huang K, Huang S, Chen G, Li X, Li S, et al. An end-to-end multi-task system of automatic lesion detection and anatomical localization in whole-body bone scintigraphy by deep learning. Bioinformatics 2023, 39(1).
Motegi K, Miyaji N, Yamashita K, Koizumi M, Terauchi T. Comparison of skeletal segmentation by deep learning-based and atlas-based segmentation in prostate cancer patients. Ann Nucl Med 2022, 36(9):834-841.
Liu S, Feng M, Qiao T, Cai H, Xu K, et al. Deep Learning for the Automatic Diagnosis and Analysis of Bone Metastasis on Bone Scintigrams. Cancer Manag Res 2022, 14:51-65.
Wuestemann J, Hupfeld S, Kupitz D, Genseke P, Schenke S, et al. Analysis of Bone Scans in Various Tumor Entities Using a Deep-Learning-Based Artificial Neural Network Algorithm-Evaluation of Diagnostic Performance. Cancers (Basel) 2020, 12(9).
Han S, Oh JS, Seo SY, Lee JJ. Performance of deep learning models for response evaluation on whole-body bone scans in prostate cancer. Ann Nucl Med 2023, 37(12):685-694.
Hajianfar G, Sabouri M, Salimi Y, Amini M, Bagheri , et al. Artificial intelligence-based analysis of whole-body bone scintigraphy: The quest for the optimal deep learning algorithm and comparison with human observer performance. Z Med Phys 2023.
Murata T, Hashimoto T, Onoguchi M, Shibutani T, Iimori T, et al. Verification of image quality improvement of low-count bone scintigraphy using deep learning. Radiol Phys Technol 2024, 17(1):269-279.
Ito T, Maeno T, Tsuchikame H, Shishido M, Nishi K, et al. Adapting a low-count acquisition of the bone scintigraphy using deep denoising super-resolution convolutional neural network. Phys Med 2022, 100:18-25.
Ichikawa S, Sugimori H, Ichijiri K, Yoshimura T, Nagaki A. Acquisition time reduction in pediatric (99m) Tc-DMSA planar imaging using deep learning. J Appl Clin Med Phys 2023, 24(6):e13978.
Pan Z, Qi N, Meng Q, Pan B, Feng T, et al. Fast SPECT/CT planar bone imaging enabled by deep learning enhancement. Med Phys 2024.
de Hond AAH, Steyerberg EW, van Calster B. Interpreting area under the receiver operating characteristic curve. Lancet Digit Health 2022, 4(12):e853-e855.
Bahloul A, Verger A, Lamash Y, Roth N, Dari D, et al. Ultra-fast whole-body bone tomoscintigraphies achieved with a high-sensitivity 360 degrees CZT camera and a dedicated deep-learning noise reduction algorithm. Eur J Nucl Med Mol Imaging 2024, 51(5):1215-1220.
Ritt P. Recent Developments in SPECT/CT. Semin Nucl Med 2022, 52(3):276-285.
Minarik D, Enqvist O, Tragardh E. Denoising of Scintillation Camera Images Using a Deep Convolutional Neural Network: A Monte Carlo Simulation Approach. J Nucl Med 2020, 61(2):298-303.
Ardenfors O, Svanholm U, Jacobsson H, Sandqvist P, Gryback P, et al. Reduced acquisition times in whole body bone scintigraphy using a noise-reducing Pixon(R)-algorithm-a qualitative evaluation study. EJNMMI Res 2015, 5(1):48.
Qi N, Pan B, Meng Q, Yang Y, Feng T, et al. Deep learning enhanced ultra-fast SPECT/CT bone scan in patients with suspected malignancy: quantitative assessment and clinical performance. Phys Med Biol 2023, 68(13).
Pan B, Qi N, Meng Q, Wang J, Peng S, et al. Ultra high speed SPECT bone imaging enabled by a deep learning enhancement method: a proof of concept. EJNMMI Phys 2022, 9(1):43.
Dickson JC, Armstrong IS, Gabina PM, Denis-Bacelar AM, Krizsan AK, et al. EANM practice guideline for quantitative SPECT-CT. Eur J Nucl Med Mol Imaging 2023, 50(4):980-995.
Halim F, Yahya H, Jaafar KN, Mansor S. Accuracy Assessment of SUV Measurements in SPECT/CT: A Phantom Study. J Nucl Med Technol 2021, 49(3):250-255.
Kaneta T, Ogawa M, Daisaki H, Nawata S, Yoshida K, et al. SUV measurement of normal vertebrae using SPECT/CT with Tc-99m methylene diphosphonate. Am J Nucl Med Mol Imaging 2016, 6(5):262-268.
Beck M, Sanders JC, Ritt P, Reinfelder J, Kuwert T. Longitudinal analysis of bone metabolism using SPECT/CT and (99m)Tc-diphosphono-propanedicarboxylic acid: comparison of visual and quantitative analysis. EJNMMI Res 2016, 6(1):60.
Arvola S, Jambor I, Kuisma A, Kemppainen J, Kajander S, et al. Comparison of standardized uptake values between (99m)Tc-HDP SPECT/CT and (18)F-NaF PET/CT in bone metastases of breast and prostate cancer. EJNMMI Res 2019, 9(1):6.
Qi N, Meng Q, You Z, Chen H, Shou Y, et al. Standardized uptake values of (99m)Tc-MDP in normal vertebrae assessed using quantitative SPECT/CT for differentiation diagnosis of benign and malignant bone lesions. BMC Med Imaging 2021, 21(1):39.

No competing interests reported.

Download PDF

Editorial decision: Revision requested
01 Aug, 2024
Reviews received at journal
31 Jul, 2024
Reviews received at journal
29 Jul, 2024
Reviewers agreed at journal
29 Jul, 2024
Reviewers agreed at journal
29 Jul, 2024
Reviewers invited by journal
28 Jul, 2024
Editor invited by journal
28 Jul, 2024
Editor assigned by journal
22 Jul, 2024
Submission checks completed at journal
22 Jul, 2024
First submitted to journal
10 Jul, 2024

You are reading this latest preprint version

Clinical performance of deep learning enhanced ultra-fast whole-body scintigraphy in patients with suspected malignancy

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

1 Introduction

2 Material and Methods

2.1 Patients

2.2 Image acquisition

2.3 Deep learning workflow

2.4 Image analysis

2.4.1 Subjective assessment

2.4.2 Objective assessment

2.5 Statistical analysis

3 Results

4 Discussion

5 Conclusion

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1