The research study identified 55 related publications that qualify for reporting. As illustrated in Fig. 3, there has been an increasing number of publications over the last 5 years. Starting from one publication in 2017, this number reached 17 in 2021.

Fig. 3

Five major application domains were identified during each qualified study's analysis: diagnosis/classification, denoising, full-count SPECT estimation, SPECT attenuation correction (AC), and reconstruction. The latter four domains belong to a broader spectrum: image quality improvement. The distribution of publications among these domains is presented in Fig. 4.

Fig. 4

The study reviewed publications presented in scientific journals and conferences. Out of the 55 reviewed publications, 45 are published in peer-reviewed journals, as illustrated in Fig. 5. An analytical overview of the journals and conferences that participate in publishing is presented in Fig. 6.

Fig. 5

Number of publications in scientific journals and conferences

Fig. 6

Overview of the publishing journals and conferences

Diagnosis/classification

Cardiovascular disease diagnosis was the most popular domain of DL application in SPECT studies. It refers to CAD diagnosis, myocardial defect identification, and abnormality detection in Polar Maps or SPECT scans. CNNs are the dominant DL strategy for this type of classification. Several research studies deploy state-of-the-art pre-trained CNNs that have already succeeded in relevant applications. In contrast, many studies develop their own CNN architectures intending to propose task-specific models that seek and extract medical image features. ML methods have also been evaluated in recent studies. ML methods differentiate themselves from the DL methods because they do not process the SPECT images directly. Instead, they analyze clinical data or predefined image features for the same task. Table 1 summarizes the presented literature regarding the diagnosis and classification tasks. Table 2 showcases studies that perform external validation of the proposed techniques.

Table 1 List of identified diagnosis/classification studies along with their main characteristicsTable 2 Classification studies performing external validationHandcrafted CNNs

Various studies propose handcrafted CNNs that distinguish between normal and abnormal SPECT images visualizing the myocardia, or Polar Maps, which summarize the 3D information of multiple heart views into a single polar plot.

Papandrianos et al. [18] presented an RGB-CNN model to classify SPECT images concerning their abnormal findings. There are a total of 513 cases, and they are represented in stress and rest conditions. The problem that is addressed is the differentiation of normal and ischemic images. Data scarcity issues were circumvented by applying data augmentations. The proposed model accomplished 90.2% accuracy and a 93.77% AUC value in discriminating ischemic from normal SPECT images, with the human reader interpretation considered as the ground truth. These results demonstrate the magnificent capability of the model to predict correctly, despite the small dataset. The same research team [19] extended their previous work to diagnose ischemia and/or infarction using CNNs. The dataset of the corresponding research includes a total of 224 patients who had undergone stress and rest SPECT tests. The participants underwent invasive coronary angiography (ICA) 40 days after MPI. Two DL techniques were followed: an (a) implementation of RGB-CNN from scratch and (b) transfer learning to classify images as normal or abnormal. The pre-trained models were VGG16, DenseNet, MobileNet, and InceptionV3. With reference to the visual assessment performed by medical experts, the results reported significant abilities of the proposed CNN with an overall accuracy of 93.48 ± 2.81%. This accuracy is significantly improved compared to the 90.2%, which was initially obtained in [18].

Narges Zahiri et al. [20] aimed to explore the potential of deep CNNs to distinguish between normal and abnormal polar maps with reference to the physician’s diagnosis. The dataset included 3318 stress and rest polar maps. Data augmentation was utilized to expand the training dataset. The proposed DL model was thoroughly validated under a fivefold cross-validation procedure. The model achieved a 0.845 AUC. Besides, the inclusion of rest perfusion maps significantly improved the AUC of the DL model (AUC: 0.845) compared with stress polar maps only (AUC: 0.827). Papandrianos et al. [21] explored the potential of automatic classification of polar maps between normal and abnormal by implementing a custom RGB CNN. The study included 314 polar maps in stress, rest representation, and AC and NAC formats. RGB-CNN was trained using physician interpretation as ground truth. The RGB-CNN proposal competed against the pretrained VGG-16 network. According to the results, RGB extracted 92.07% and VGG-16 95.83%. RGB-CNN competed against robust state-of-the-art methods.

Some research compares DL-based results against quantifiable metrics advised by the guidelines. For example, Yuka Otaki et al. [22] developed a DL model to identify CAD and compared its results against the Total Perfusion Deficit (TPD) method. One thousand one hundred sixty patients were included to classify raw upright and supine stress MPI polar maps. MPI and ICA were performed within a 6-month interval. As an external validation method, leave-one-centre-out was utilized with four models. Julian Betancur et al. [23] designed a CNN for the same purpose. The number of participants was 1160, whilst the utilized data involved semi-upright and supine stress Polar Map representations. The classification of obstructive disease was evaluated using the leave-one-centre-out cross-validation technique with four centres, where all validated predictions were merged to avoid a calculation for a single centre. The CNN model performs the diagnosis without adding predefined coronary territories. The performance of CNN was compared against combined perfusion quantification by TPD, achieving 84.8% sensitivity versus 82.6% obtained with clinical reading. In a subsequent study, Julian Betancur et al. [24] evaluated the automatic diagnosis of CAD from SPECT image inputs in contrast with TPD with a deep CNN. A total of 1638 patients without known CAD and with ICA performed within 6 months of MPI were examined. The data involved raw and quantitative polar maps in only stress representation. A stratified tenfold cross-validation procedure was adopted. The AUC score for disease prediction by their proposed DL scheme was superior to TPD (per patient: 0.80 vs 0.78; per vessel: 0.76 vs 0.73). With the DL threshold set to the same specificity as TPD, per-patient sensitivity improved from 79.8% (TPD) to 82.3%, and per-vessel sensitivity improved from 64.4% (TPD) to 69.8%.

Besides distinguishing between normal and abnormal subjects, some studies aim to perform region-based classification. Arvidsson et al. [25] developed a CNN to predict obstructive coronary artery disease in the left anterior artery, left circumflex artery, and right coronary artery using SPECT Polar Maps. A total of 588 patients were included in this study, whilst clinical data like angina symptoms and age were also utilized. The proposed CNN framework achieved an average AUC of 0.89 per vessel and 0.95 per patient, using the ICA findings as a reference. Furthermore, gradient-weighted class activation mapping (Grad-CAM) was utilized to visually demonstrate the regions on which predictions are based to extract the output. The authors observed sex differences in the diagnostic performance of DL for the prediction of obstructive CAD from D-SPECT, with DL outperforming visual and TPD in men but not in women.

An increasing number of works propose explainable DL-based methods that perform image classification and inform the user about the suggested areas of interest wherein the model bases its predictions. Miller et al. [26] utilized an explainable DL model to improve the diagnostic accuracy of CAD and aid physical interpretation. A total of 240 patients underwent MPI examinations and were included in this study, with ICA as a reference. Regarding the results, human readers using the DL’s prediction achieved an AUC of 0.779, whereas their interpretation without DL reached an AUC of 0.747. It is worth mentioning that DL, on its own, achieved an AUC of 0.793. Yuka Otaki et al. [27, 28] proposed an explainable DL model to detect obstructive CAD. A total of 3578 patients with suspected CAD from 9 centres were enrolled. The authors proposed a hand-crafted CNN to process the SPECT Polar Maps in stress conditions. In the fully-connected layer of the CNN, the authors supplied the sex and age of the patient to increase the number of features. Concerning ICA findings, this method achieved an Area Under Curve (AUC) score of 0.83 following a tenfold cross-validation procedure, which was superior to the quantitative analysis results by expert readers (AUC = 0.8). Also, attention maps were produced to highlight the regions and segments contributing most to the per-vessel prediction.

Singh et al. [29] developed an explainable deep learning model to predict nonfatal myocardial infarction (MI) or death, which also provides highlighted image regions related to obstructive CAD. The study included 20,401 patients, who went under SPECT MPI procedure for training and internal testing purposes and 9019 patients were added from external testing group gathered from two different sites. The external testing group was included to evaluate generalizability. The dataset consisted of polar maps in stress and rest representation with the inclusion of age, sex and cardiac volumes, which were added at the first fully connected layer. Referring to explainability, Grad-CAM was developed. For comparison reasons a logistic regression model was developed with the following values age, sex, stress total perfusion deficit (TPD), rest TPD, stress left ventricular ejection fraction, and stress left ventricular end-systolic volume. The model achieved and AUC of 0.76, which is higher than stress TPD with 0.63 AUC, ischemic TPD with 0.6 AUC and compared to logistic regression model, which extracted 0.72 AUC. The developed model improved accuracy in contrast to traditional quantitative approaches and is well calibrated and provides robust results.

Jui-Jen Chen et al. [30] examined 979 SPECT subjects from a local hospital for the diagnosis. However, whether the images are labelled based on experts' visual inspection or on the ICA's findings is not reported. A three-dimensional CNN has been applied to classify the SPECT slices. Furthermore, Grad-CAM heat maps have been produced to identify myocardial defects in the images. The proposed model obtained accuracy, sensitivity, and specificity metrics of 87.64%, 81.58%, and 92.16%, respectively, in distinguishing between normal and abnormal images using a test set of 89 samples. Nathalia Spier et al. [31] investigated Graph CNNs for CAD diagnosis. They enrolled 946 polar map images in stress and rest representations of the heart. Labelling has been done using the human observer interpretation. Also, heatmaps were produced and demonstrated the segments of the heart that were indicated as pathological. The extracted results demonstrate adequate performance in classifying unseen data under a fourfold cross-validation procedure, in contrast to clinical visual analysis, with 92.8% and 95.9% specificity in rest and stress data, respectively. The proposed model achieves an agreement with the human observer on 89.3% of rest test polar maps and on 91.1% of stress test polar maps. Localization performed on a fine 17-segment division of the polar map achieves an agreement of 83.1% with the human observer.

CNNs and transfer learning

Selcan Kaplan Berkaya et al. [32] intended to produce a classification model to classify SPECT images and identify perfusion abnormalities like ischemia and infarction. The summed stress and rest images from 192 patients were studied. Two models were proposed. The first is a DL-based model which employs State Of The Art (SOTA) CNNs and fully-connected layers of support vector machines (SVM) for the classification of the deep extracted image features. As far as the second model is concerned, it involves image processing techniques like segmentation, feature extraction, and colour thresholding applied to segmented parts of each SPECT slice. This method extracts five predefined image features classified by a rule-based algorithm. With reference to the visual assessment as performed by the experts, the integrated CNN-SVM model achieved 92% accuracy, 84% sensitivity, and 100% specificity, whereas the knowledge-based classification attained 93% accuracy, 100% sensitivity, and 86% specificity. Those metrics are reported on a test dataset that includes 17% of the total samples. Hui Liu et al. [33] demonstrated a DL approach to automatically diagnose myocardial perfusion abnormalities in abnormal and normal with only stress MPI profile maps as input. A total of 37,243 patients who underwent stress-only and stress/rest SPECT MPI have been examined. The study involved three SPECT/CT cameras. There was an addition of six extra features, including gender, BMI, length, stress type, radiotracer, and the option of including or not including the attenuation correction. The ResNet-34 model is employed to perform the feature extraction. The results were compared against the conventional quantitative perfusion defect size (DS) method. With reference to the diagnostic impression from nuclear cardiologists, the model achieved an AUC of 0.87, outperforming the DS method. Also, the proposed network showed robustness to image acquisition device variation, achieving an 82% and 84% accuracy in all scanners. The model also achieved greater performance in female participants, reaching an accuracy of 87%.

Apostolopoulos et al. [34] used the Polar Map images under stress and at rest to diagnose CAD using the pretrained VGG16 model. The study involved 216 participants. The attenuation correction (AC) and the non-attenuation correction (NAC) Polar Map images were merged into a single image per patient. With reference to the findings of ICA, VGG16 achieved an accuracy of 74.53%, a sensitivity of 75.00% and a specificity of 73.43%. The respective figures for MPI interpretation by experienced nuclear medicine physicians were 75.00%, 76.97%, and 70.31%. The accuracy of semi-quantitative polar map analysis was lower, at 66.20% and 64.81% for the AC and NAC techniques, respectively. Besides, the model showed robustness to acquisition device variation. The same author team extended their study [35] by proposing a hybrid CNN-Random Forest approach for classifying Polar Map images and clinical attributes into normal and abnormal classes, using ICA findings as a reference for CAD disease. The study involved 566 patient cases. The authors used the InceptionV3 pretrained model to predict the class of the input Polar Maps. The model's output was considered a unique attribute among 22 clinical factors, such as gender and age. The Random Forest classifier was employed to predict the outcome. With reference to ICA results, the model achieved 78.44% accuracy, 77.36 sensitivity, and 79.25% specificity. The human cognitive process's overall accuracy reached 79.15%, which is approximately 1% better than the automatic model's accuracy (78.43%). Besides, the overall agreement rating between the human experts and the model was 86% (Cohen’s Kappa = 72.24). The model was also tested on unseen data from a different SPECT scanner and achieved consistent results (76.53% accuracy).

Trung et al. [36] proposed a CNN to diagnose CAD. The authors utilized polar maps and SPECT slices. The dataset included 1413 heart SPECT images labelled by a nuclear expert as CAD and non-CAD. The DL network’s (VGG-16) performance was evaluated using fivefold cross-validation. The results indicated that SPECT images guarantee a better diagnosis than polar maps, with a precision of 86.14% ± 2.14% and 82.57% ± 2.33%, respectively.

Machine learning

Kenichi Nakajima et al. [37] proposed an ANN for CAD diagnosis and myocardial ischemia and/or infarction detection. This research consisted of 1001 stress/rest MPI images for training, and there was an addition of 364 images for validation. Expert interpretations served as the gold standard. The achieved results were compared against the conventional quantitative approach. The ANN algorithm outperformed the conventional summed difference scores, scoring an AUC of 0.92 in identifying stress defects and 0.91 in stress-induced ischemia.

Souza Filho et al. [38] explored the potential of developing different ML models like Adaptive Boosting (AB), Gradient Boosting (BG), Random Forest (RF), and Extreme Gradient Boosting (XGB) to find the ideal model for efficient differentiation between normal and abnormal cases of SPECT Polar Maps labelled by human readers. The stress and rest conditions included a total of 1007 Polar Maps. Each image was divided into five horizontal and five vertical slices, where the sum of pixel intensities from each slice was computed, and ten attributes were acquired. Afterwards, data augmentation was applied to generate 324 Polar Maps. RF was concluded to have the best sensitivity with 96%, whereas AB, GB, and XGB obtained 92%, 94%, and 95%, respectively.

Hu et al. [37] evaluated the per-vessel and per-patient predictions using an ML methodology. A total of 1980 patients were utilized in stress and rest demonstrations, and overall, 18 clinical, nine stress test and 28 imaging variables were utilized for this study. The model achieved an AUC of 0.79 on a patient-level basis and 0.81 on a vessel-level basis using ICA findings for reference. Baskaran et al. [39] investigated the importance of including clinical and imaging variables for the successful prediction and revascularization of CAD by developing XGBoost and estimated the results by developing a fivefold cross-validation. Seven hundred nineteen ICA-confirmed patients were included in this research. The proposed model performed similarly to previous history-based scores and achieved an AUC of 0.779, a sensitivity of 89.2%, and a specificity of 92.9%. Following the results, BMI is the most valued non-imaging variable to be included in the prediction and revascularization of CAD. Nevertheless, BMI, age, and angina severity are the most important parameters for prediction.

Betancur et al. [40] evaluated the inclusion of clinical and SPECT MPI data to predict MACE (Major Adverse Cardiac Events) by developing an ensemble boosting algorithm, LogitBoost. A total of 2617 patients were considered under stress examination. Twenty-eight clinical variables, seventeen stress test variables, and twenty-five imaging variables were included. Furthermore, LogitBoost with both clinical and imaging data (ML-Combined) was compared against the utilization of only imaging variables as input, and visual diagnosis and automated quantitative imaging analysis, and ML-combined outperformed with an AUC of 0.81. Rahmani et al. [41] aimed to investigate the integration of ANN to predict obstructive CAD by adding clinical data. Ninety-three polar maps were included, with the patients in stress and rest demonstrations. Regarding the clinical data, various combinations were examined, and the accuracy increased with age, gender, and the number of cardiac risk factor additions. ANN achieved 85.7% accuracy and improved the results by adding patient data.

Image quality improvement

Image quality improvement alludes to various improvements, including low-count SPECT estimation, AC, de-noising, and reconstruction. Table 3 presents the reviewed literature.

Table 3 List of identified image quality improvement studies along with their main characteristicsLow-count SPECT image estimation

Reduction of human body radiation exposure is highly desirable. Reducing radiation exposure involves a low-dose SPECT scan with low-count emission data. Full-count SPECT image outcome is difficult to estimate from low-count data based on the existing image processing and de-noising methods. Besides, the low-count SPECT noise is different from the full-count noise. Fast-scan is another way to reduce radiation exposure and the patients’ discomfort during the examination. Patient pain and discomfort are responsible for artefacts due to motion. Fast scan results in low-count SPECT images. DL methods address such issues and estimate the full-count SPECT scans given the low-dose or fast-scan outcome.

A.

Low-dose

Ramon et al. [42] proposed a 3D CNN based on CAEs to estimate the standard-dose SPECT image from the low-dose image. The study included 930 SPECT scans simulated at 1/8 and 1/16 of the standard clinical dose. The authors evaluated their method using the average correlation between the estimated and standard dose images. Also, the estimated images were compared to those obtained from conventional image de-noising methods (spatial post-filtering). When estimating the standard dose from 1/16 dose, the proposed method achieved similar image quality to the quality obtained from \( \!\mathord}\right.\kern-0pt} \!\lower0.7ex\hbox}\) dose with conventional de-noising. In another study by Olia et al. [43], the authors explored the results of predicting the standard-dose image from a low-dose setup at half, quarter, and one-eighth dose levels. The study involved 345 patients. A GAN architecture was deployed to decrease the administered activity, ensuring stable accuracy and clinical values to achieve the standard dose image estimation. With reference to the actual standard dose images, the highest PSNR and SSIM and lowest RMSE were attained at a half-dose level. Overall, the proposed network can increase the quality of high low-dose SPECT images with 100% acceptance, according to a nuclear medicine specialist.

Ramon et al. [44] investigated the application of different 3D DL methodologies to suppress noise in low-dose SPECT MPI images. The dataset includes 1052 patients, and two reconstruction methods were applied, namely FBP (Filtered Back-Propagation) and OSEM (Ordered-Subsets Expectation–Maximization). The authors trained the model with low-dose acquisitions as input and full-dose images as target and explored different numbers of dose levels (1/2, ¼, 1/8 and 1/16 of full dose). Reviewing the results, the proposed DL approach can reduce substantial noise and enhance the accuracy compared to conventional reconstruction filtering. More specifically, with ½ dose, the model achieved 0.799 AUC, whereas full dose attained 0.801 AUC.

Similarly, Song et al. [45] explored a de-noising methodology based on a three-dimensional residual CNN for low-dose cardiac-gated SPECT images. The study includes 119 clinical cases in total. Regarding the model's training, the CNN utilized as a training dataset included the low-dose images with a 25% reduction of radiation dose as input and the corresponding full-dose images as output. The proposed CNN methodology was compared against traditional methods based on the ST-NLM (SpatioTemporal Non-Local Means) technique, and CNN attained an nMSE of 0.153, where ST-NLM and Gaussian post-filter extracted 0.163 and 0.172. Furthermore, the CNN decreased the nMSE by 6.13% and the ST-NLM, Gaussian post-filter, reduced it by 6.13%. Overall, the proposed CNN enhanced the noise reduction in the reconstructed myocardium and the spatial resolution of the LV wall.

Song et al. [46] developed a spatiotemporal CNN (ST-CNN) model for image denoising in low-dose cardiac gated SPECT studies. A total of 119 cases were included, and the proposed model is trained with low-dose images as input and includes full-dose images as output. Moreover, the authors included in the developed model an LSTM component in order to perform correctly with the format of a gated sequence. The corresponding model was compared against spatial-only S-CNN and ML reconstruction, where ST-CNN outperformed with NMSE 0.127, and S-CNN and ML extracted 0.161 and 0.273, respectively.

B.

Fast-scan

Estimating the standard acquisition time image given a fast scan is seldom investigated in the literature. In the only work discovered, Shiri et al. [17] aimed to reduce the acquisition time of acquiring SPECT images from patients by exploring two approaches. The first approach refers to the reduction of scanning time per projection. The second approach refers to reducing the number of acquired projection images during acquisition. The study includes 363 cases with normal patients but various heart disorders, like infarction and ischemia, where the SPECT data were reconstructed with the OSEM algorithm. For each patient, four datasets were produced: FT (full-time projections), HT (half-time acquisition per projection), FP (full projections) and HP (half projections). The proposed method is applying a residual network, namely ResNet, to predict FT from HT and FP from HP images, and the results were evaluated with tenfold cross-validation. According to the results, the predicted FT had better image quality than the predicted FP, with a decreasing RMSE of 8.0 ± 3.6 and 6.8 ± 2.7 for FT and FP, respectively.

Moreover, the HP reconstructed images acquired better quality than the HT reconstructed images. The error increases as acquisition time is reduced. The deep neural network can effectively restore image quality.

Attenuation correction

The majority of dedicated cardiac SPECT scanners do not have integrated CT technology. As a result, attenuation correction (AC) for image quantification is very challenging due to the presence of artefacts. Several research papers address this issue by introducing DL-based AC methods.

Several studies employ the U-Net CNN to estimate the NAC image directly or generate the attenuation maps that deliver the AC image. In [47], Yang et al. used a Deep CNN to generate the attenuation-corrected SPECT from the NAC scan. The study involved 100 participants. The effectiveness of the proposed method was verified by voxelwise and segment-wise analyses against the reference, CT-based AC using the 17-segment myocardial model of the American Heart Association under a tenfold cross-validation procedure. Voxelwise correlations with the reference image were 97.7% ± 1.8% (slope, 0.94; R2 = 0.91), whereas the segmental errors stayed mostly within ± 10%. The generated Polar Maps were visually assessed for artefact reduction. The study showed promising results, but the performance of the proposed method was affected by the amounts of attenuation introduced between the scans and the different observed uptake patterns. In another work, Mostafapour et al. [48] analyzed the direct attenuation correction of SPECT MPI images, utilizing two DL-based algorithms, ResNet and U-Net. The dataset consisted of 99 patients, including both normal and abnormal cases. Moreover, the Chang AC approach [49] was applied for comparison against DL models. Based on the quantitative metrics and external evaluation of 19 images, the DL approaches produced images that agree with SPECT CT-AC images, whereas the Chang approach underestimated the patient’s status based on the horizontal profile. ResNet and U-Net achieved a ME of 6.99 ± 16.72 and − 4.41 ± 11.8 and an SSIM of 0.99 ± 0.04 and 0.98 ± 0.05, respectively. The Chang approach extracted the ME and SSIM of 25.52 ± 33.98 and 0.93 ± 0.09, respectively.

Mostafapour et al. [50] investigated the generated attenuation-corrected images utilizing ResNet and U-Net. This research enrolled 99 patient cases. NAC SPECT images were included as input, and CT-based attenuation-corrected images were used as reference. Nineteen cases were provided as an external validation dataset to further evaluate the models. Chang’s method [49] was compared against the deployed ResNet and U-Net approaches and was found inferior, with a ME of − 6.99 ± 16.72, against − 4.41 ± 11.8, and an SSI of 0.99 ± 0.04 and 0.98 ± 0.05, respectively. Chen et al. [51] explored the capabilities of transfer learning and utilized the state-of-the-art networks U-Net and DuRDN to generate attenuation maps from SPECT. A total of 200 SPECT/CT cases were included in this research. Regarding the results, DuRDN outperformed the prediction of μ-maps and the reconstruction of SPECT AC images. The concluded error between ground-truth and predicted μ-maps is 5.13 ± 7.02 and between ground-truth and reconstructed images is 1.11 ± 1.57%.

Chen et al. [52] compared the efficiency between direct and indirect techniques for dedicated SPECT and general purpose SPECT datasets by developing U-Net and DuRDN. In both approaches, AC was performed using CT-derived μ-maps as ground truth. More specifically, in indirect methodologies, attenuation maps (μ-maps) are generated from emission images, whereas in direct methodologies, attenuation-corrected (AC) images are predicted directly from non-attenuation (NAC) images without the need for μ-maps. Concerning general purpose SPECT, the study involved 400 participants, who underwent stress and rest examinations, where in the direct approach, both photopeak and scatter images were concatenated and inserted as input into the networks to predict the corresponding AC images directly. In indirect methodologies with general purpose SPECT, the NAC images were first concatenated. They were then applied as input to the networks to predict the intermedia μ-maps, which were then utilized for the iterative reconstructions to output the predicted AC images. The indirect strategies with DuRDN as a DL approach for dedicated SPECT with full μ-maps achieved better results with an nMSE (average normalized Mean Squared Error) of 1.2 ± 0.72%, in contrast to the 2.21 ± 1.17% yielded by past direct methodologies. Overall, for both SPECT systems, the indirect approaches demonstrate stability and efficiency in contrast to direct approaches, where the direct image-to-image transformation might not ensure constancy.

GANs enjoy remarkable success in generating the AC image directly from the NAC input. For example, Shi et al. [53] developed a 3D CNN based on a cGAN framework to estimate attenuation maps directly from emission data. Sixty-five cardiac SPECT images were included, where both photopeak and scatter were inserted. Clinical characteristics such as gender, age, height, weight, and BMI were also incorporated. The corresponding patients went on a 1-day stress-only low-dose protocol. The proposed model achieved an nMAE of 3.6%0.85% on a test set of 25 images, ensuring that the model can generate trustworthy attenuation maps consistent with CT-based maps. Liu et al. [54] explored the potential of the PRAC (Post-Reconstruction Attenuation Correction) approach combined with DL methodology to provide accurate AC images for SPECT systems. The study included 30 SPECT clinical cases in stress demonstration. The researchers developed a 3D GAN model to synthesize the attenuation map directly from the NAC SPECT image. Following this, the PRAC image was reconstructed utilizing the synthesized map and the virtual projections. For further evaluation, the PRAC image was generated based on the DL attenuation map and the CT-based attenuation map. The results were compared with scanner-generated AC images to serve as the reference ground truth. Following the post-reconstruction AC, both approaches performed consistently in contrast with scanner-generated NC images. Overall, the PRAC method with both approaches can enhance the correlation with the scanner-generated AC images compared with scanner-generated NC images. Furthermore, PRAC-CT outperformed PRAC-DL regarding scattering. In terms of metrics, the PRAC-CT extracted SSIM of 0.946 ± 0.041 compared to the PRAC-DL of 0.902 ± 0.056. However, both methodologies reduce ROI biases after attenuation correction. The authors developed a PRAC approach based on scanner-generated NC images without adding raw data from CT-less attenuation correction.

Shanbhag et al. [55] developed a conditional generative adversarial neural network model to generate simulated AC images straightly from NAC images, without including the use of CT. The dataset included 4886 patients, where short-axis NC and AC images are demonstrated for training purposes and 604 patients from two separate external sites included for testing purposes. For comparison reasons the authors gathered the results of stress TPD attained from NC, AC and DeepAC (generated of the proposed model) images. The proposed model achieved 0.79 AUC compared to NC TPD with 0.7 AUC and similar with AC TPD. With respect to normalcy rate the generated simulated images produced better results with 70.4% and 75.0% for DeepAC TPD and AC TPD accordingly, in contrast with NC TPD which extracted 54.6%. As a conclusion, the developed model enhanced the diagnostic accuracy for obstructive CAD and can function without the need of CT hardware and produced results similar to actual AC images.

Hagio et al. [56] proposed a convolutional neural network based on deep learning methodology to generate “virtual” attenuation-corrected polar maps from NAC data, without adding CT imaging scans. The study includes 11,532 cases with paired NAC and CTAC images. The authors developed a DL algorithm based on the U-Net architecture framework to predict DLAC polar maps from NAC polar maps. The produced model attained 0.827 AUC, in contrast with NAC images, which extracted 0.78 AUC. Regarding sensitivity and specificity, the produced model extracted 88% sensitivity and achieved 18.9% increased value of specificity for DLAC and 25.6% for CTAC polar maps. Conclusively the developed model generated similar attenuation-correction images with CTAC and accomplished better diagnostic accuracy with exceptional overall performance.

Similar results have been reported in other studies as well [