Examination of simple artificial intelligence-based analysis of dopamine transporter scintigraphy for supporting a diagnosis of Parkinson’s disease

doi:10.21203/rs.3.rs-7387754/v1

Download PDF

Research Article

Examination of simple artificial intelligence-based analysis of dopamine transporter scintigraphy for supporting a diagnosis of Parkinson’s disease

https://doi.org/10.21203/rs.3.rs-7387754/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 26 Nov, 2025

Read the published version in Annals of Nuclear Medicine →

You are reading this latest preprint version

Introduction:

In the Movement Disorder Society criteria for the diagnosis of Parkinson's disease (PD), evaluation of the presynaptic dopamine system should be performed using dopamine transporter single-photon emission computed tomography (DAT SPECT). However, it is difficult for unexperienced physicians to detect a mild defect. Here, we attempted to develop a simple deep learning-based image analysis method to evaluate DAT SPECT images.

Methods

We used data from 300 patients who were diagnosed with PD and 102 control patients with non-neurodegenerative diseases as the artificial intelligence (AI) development cohort. For validation, we analyzed the data of 96 patients with PD from an independent cohort. We divided the development cohort into the training and test sets. Using the training set, we performed transfer learning using six pre-trained convolutional neural network architectures, and created AI models. We evaluated their accuracy, sensitivity, and area under the receiver operating characteristic curve, and further confirmed their validity by using the validation cohort. In addition, we compared the accuracy of the best AI model with that of two experienced neurologists and a resident.

Results

The selected AI model could interpret DAT SPECT images with an accuracy of 0.959; accuracy in the validation cohort was 0.8959–1. There was no significant difference between the accuracy of the AI model and physicians.

Conclusion

Our simple AI model for the interpretation of DAT SPECT images was accurate and robust. Its accuracy was equivalent to that of physicians.

Dopamine transporter scintigraphy

Parkinson’s disease

AI model

Deep learning

Convolutional neural network

Parkinson's disease (PD) is a progressive degenerative disorder that is mainly characterized by the degeneration of nigrostriatal dopamine neurons and presents with various symptoms such as constipation, rapid eye movement sleep behavior disorder, olfactory dysfunction, slowness of movement, tremor, and cognitive dysfunction [1]. The incidence of PD increases rapidly with age, and as the world's population is also aging, the number of patients is increasing significantly [2]. Despite this situation, there are a limited number of physicians who are sufficiently trained to make a diagnosis of PD, and it is predicted that the opportunities for general physicians to encounter PD will increase in the future. Thus, there is an urgent need to develop a simple diagnostic tool that will enable the diagnosis of PD at the primary care level.

In recent years, artificial intelligence (AI) has emerged and begun to be adopted gradually in medical imaging, data analysis, diagnosis, and healthcare [3]. As a recent example, the interpretation of radiological images for the differentiation between coronavirus disease 2019 pneumonia and general pneumonia has been well reported, and in fact, a highly accurate system has been developed [4]. AI-based image analysis is also being developed for interpreting dopamine transporter single-photon emission computed tomography (DAT SPECT) images for the diagnosis of PD [5]. This method involves machine learning to extract, calculate, and analyze features such as image area and length to determine abnormalities and diagnose PD. Reports indicate that the combination of calculated striatal circularity and the specific binding ratio (SBR) is useful for distinguishing PD from normal conditions [6]; that multivariate analysis combining uptake intensity, left-right asymmetry, morphology, and patient age is superior [7]; and that image analysis using a support vector machine is effective [8]. These methods require humans to set features and, while they are highly effective in accurately collecting and analyzing known information and features, further improvements are limited.

Recently, new analytical models using deep learning, in which the AI itself learns from the images and detects features, have begun to be applied for DAT SPECT analysis. Examples include a new three-dimensional (3D) image analysis model [9, 10], a method that selects specific slices, aligns them, and interprets them as two-dimensional (2D) images [11], a method that extracts one image from all results and analyzes it [12], and a method that processes and analyzes key images [13, 14]. However, the number of such reports is still limited. Furthermore, creating new 3D image analysis models requires high-performance computers and highly specialized knowledge of AI technology, as well as addressing issues such as the instability caused by using low-resolution DAT SPECT images and the discrepancy between the precision of the analysis equipment. In 2D image analysis, in addition to the loss of 3D information, there are challenges such as selecting target images when extracting single or multiple images and the processing methods for multiple-image models.

Here, we aimed to develop a simple deep learning-based image analysis method to evaluate dopaminergic neuron loss in DAT SPECT images using an easy-to-access, pre-trained convolutional neural network (CNN).

Study population

The overall flow of this study is shown in Fig. 1. For the development of the AI model, we used data from the electronic medical records of Nagoya University. There were 1,163 consecutive cases who visited the Department of Neurology at Nagoya University Hospital and underwent DAT SPECT scans between May 2014 and April 2021. Of these, 357 cases were clinically diagnosed with PD, whereas 300 cases were diagnosed as “clinically established” or “clinically probable” PD according to the International Parkinson and Movement Disorder Society diagnostic criteria [15]. We defined these 300 patients as the PD group. We also analyzed the data of 102 cases diagnosed with non-neurodegenerative diseases as the Control group. We then divided the groups randomly into the training and test datasets. From the PD group, we used 209 cases as the training dataset and 91 cases as the test dataset; from the Control group, we used 71 cases as the training dataset and 31 cases as the test dataset. For external validation, we used DAT SPECT data from an independent cohort at the NHO Higashinagoya National Hospital. Among 282 consecutive cases of PD who visited the hospital from May 2019 to April 2022, we analyzed data from 96 cases who fulfilled “clinically established” or “clinically probable” PD according to the International Parkinson and Movement Disorder Society diagnostic criteria.

The study conformed to the Ethical Guidelines for Medical and Health Research Involving Human Subjects endorsed by the Japanese government and was approved by Nagoya University Ethical Committee (2022 − 0362).

SPECT imaging

In the development cohort at Nagoya University, DAT SPECT images were acquired using a Symbia T (Siemens Healthineers, Erlangen, Germany) at 3–4 h after the intravenous administration of 167 MBq of ¹²³I-ioflupane (DaTSCAN; Nihon Medi-Physics, Tokyo, Japan). DAT SPECT images were acquired using the following parameter [16, 17]: low-medium energy general purpose collimation, a 128 × 128 matrix (zoom factor, 1.45), giving a pixel size of 3.3 mm and acquisition time of 28 min. DAT SPECT images were reconstructed using ordered subset expectation maximization with depth-dependent 3D resolution recovery (6 subsets, 8 iterations, 6 mm full width at half maximum Gaussian filter) with CT-based attenuation correction and energy window-based scatter correction (20% main window centered at 159 keV, 8% sub-window to touch the lower and upper sections of the main window).

For the validation cohort at NHO Higashinagoya National Hospital, at 4 h after the intravenous administration of 167 MBq of ¹²³I-ioflupane, SPECT acquisition of the entire brain was performed using a dual-head gamma camera (Symbia Evo Excel; Siemens Healthineers) equipped with a low-medium energy general purpose collimator. The parameters of acquisition were as follows: rotation radius of 14 cm, circular orbit; magnification factor of 1.45; 128 × 128 matrix, pixel size 3.3 mm; 45 projections, 4.7 s/projection. DAT SPECT images were reconstructed using ordered subset expectation maximization with depth-dependent 3D resolution recovery (6 subsets, 8 iterations, 6 mm full width at half maximum Gaussian filter) with energy window-based scatter correction (20% main window centered at 159 keV, 8% sub-window to touch the lower and upper sections of the main window). At this facility, attenuation correction was not performed. The volume of interest was defined in the bilateral striatum, and the whole brain was used as the reference.

Image analysis

We created the images to be analyzed in this study with reference to the formula used to calculate the SBR by Tossici-Bolt et al. [18] (Fig. 2). To calculate the SBR, a 44-mm slice within the region of interest that covers the striatum was integrated along the z-axis and calculated as a ratio to the reference region. We used the open-source software ImageJ (National Institutes of Health, Bethesda, MD, USA) to stack the images specified as slabs in the z-axis direction and create “the summed striatum image” as a 2D image. We performed further analysis on this 2D image (32 bits, 128 × 128 pixels).

Creating a learning machine

We created an AI model using transfer learning. As pre-trained CNN architectures, we used AlexNet [19], GoogLeNet [20], ResNet-50 [21], SqueezeNet [22], ShuffleNet [23], and MobileNet-v2 [24]. Learning was carried out using MATLAB (R2021b; MathWorks, Natick, MA, USA). We provide the training algorithm for transfer learning below.

Training algorithm for transfer learning

Use Deep Network Designer - MATLAB

Import the pre-trained model each CNN with its corresponding weights.

Replace the last three layers of the Network:

-Fully connected layer (Set the 'WeightLearnRateFactor' to 10 and the 'BiasLearnRateFactor' to 10;

and set its output to 2).

-Softmax layer

-Classification layer

Training-progress settings

Solver ->sgdm.

InitialLearnRate ->0.0001.

ValidationFrequency ->50.

MaxEpoch ->12.

MinibatchSize ->56.

ExecutionEnvironment ->auto.

SequenceLength -> longest.

SequencePaddingValue ->0.

SequencePaddingDirection ->right.

L2Rengularization ->0.0001.

GradientThresholdMethod -> l2norm.

GradientThreshold ->0.1.

ValidationPatience ->Inf.

Shuffle ->every-epoch.

CheckpointPath ->blank.

LearnRateSchedule ->none.

LearnRateDropFactor ->0.1.

LearnRateDropPeriod ->10.

ResetInputNormalization ->check.

BatchNormalizationStatistics ->population.

OutputNetwork -> last iteration

Momentum ->0.9.

We referred to the output results and adjusted InitialLearnRate, ValidationFrequency, MaxEpoch, and MinibatchSize appropriately for each CNN to achieve the best results.

Confirmation of performance

We verified the accuracy, sensitivity, specificity, false positive rate, and false negative rate of the AI models using the test datasets. We also created a receiver operating characteristic curve for the models and verified its area under the curve. Furthermore, to verify external validity, we assessed the accuracy of the models using the data from the validation cohort.

In addition, three blinded physicians interpreted the original DAT SPECT images from the test datasets and we compared the accuracy of their evaluations with those of the most accurate AI model.

Statistical analysis

To clarify the bias of the model, we applied the Shapiro-Wilk test to each group, and if both samples followed a normal distribution, we used a parametric uncorrelated t-test to calculate the p-value. We considered p-values < 0.05 to be statistically significant. We performed comparisons between groups using Fisher's exact probability test to analyze categorical variables. We used kappa values to assess the reproducibility of the evaluations by physicians and the AI model.

Characteristics of the case groups used in this study

The characteristics of the patients in the Control, PD, and PD validation groups are shown in Table 1. The diagnoses of the Control group consisted of asymptomatic healthy individuals [16, 17] (n = 80) and patients diagnosed with essential tremor (n = 22). There was no significant difference in age, but there was a significant difference in the male-to-female ratio between the groups.

Table 1
Clinical characteristics of patients
	Development			Validation
	Control	PD	p – value	PD
Number	102	300		96
Age (years old)	65.1 ± 9.2	66.1 ± 10.8	0.417*	76 ± 9.9
Sex (Male/Female)	67/35	154/146	< 0.05**	57/39
^* t-test
^** Fisher's exact probability test
PD, Parkinson’s disease

Analysis of the test dataset using the AI models

The results of analysis of the test dataset using the AI models are shown in Fig. 3. All of the AI models we created showed high accuracy of ≥ 0.93, but the model with the lowest accuracy was created from AlexNet, while those with the highest accuracy were created from ResNet-18, ShuffleNet, and MobileNet-v2. The model with the highest sensitivity was created from GoogLeNet, while the one with the lowest sensitivity was created from SqueezeNet. The model with the highest specificity was created from SqueezeNet, while the one with the lowest specificity was created from GoogLeNet. The area under the curve values of the receiver operating characteristic curves were all acceptable, with the highest value obtained from the model created from GoogLeNet and the lowest from the model created from AlexNet.

Analysis of the PD validation group using the created AI models

The results of the analysis of the validation group using the AI models are shown in Table 2. The AI models showed high accuracy in detecting abnormalities in DAT SPECT images. Among them, the AI models created from ResNet-18, SqueezeNet, ShuffleNet, and MobileNet-v2 were particularly accurate.

Table 2
Results of analysis of the validation group
	AlexNet	GoogLeNet	ResNet-50	SqueezeNet	ShuffleNet	MobileNet-v2
PD prediction	88/96	86/96	95/96	96/96	94/96	96/96
Accuracy	0.9167	0.8959	0.9896	1	0.9792	1
PD, Parkinson’s disease

Comparison of the most accurate AI model and clinicians

We finally compared the results of DAT SPECT analysis between the best AI model and clinicians. The results are shown in Table 3. To this end, we used the AI model created from MobileNet-v2 as it generated the highest accuracy using our research institute's data and the validation dataset. There was no significant difference between the AI model and three neurologists (Physicians A and B, board certified neurologists with more than 10 years of clinical experience, and C, a resident) at the precision, recall and accuracy of the analysis. For further validation, the order of all images was changed randomly, and Physicians B and C evaluated them again 1 month later without knowing their first answers. In addition, the same conditions were used for the AI model. The result confirmed the absence of a significant difference between the AI model and the physicians. In the verification of reproducibility using the κ coefficient, physicians B and C showed high reproducibility, while the AI model showed complete reproducibility.

Table 3
Accuracy and reproducibility of DAT SPECT interpretation by AI model and physicians
	Precision 1	Recall 1	Accuracy 1	p – value^a	Precision 2	Recall 2	Accuracy 2	p – value^a	κ
AI model	0.9574	0.989	0.959		0.9574	0.989	0.959		1
Physician A	0.9674	0.978	0.959	1	NR	NR	NR
Physician B	0.9375	0.989	0.9426	0.769	0.9579	1	0.9672	1	0.91
Physician C	0.9192	1	0.9344	0.57	0.9778	0.967	0.959	1	0.893
NR: Data was not reported.
Precision, Recall, Accuracy 1: Results of interpretation for 122 cases in the test group, first time.
Precision, Recall, Accuracy 2: Results of interpretation for 122 cases in the test group, one month later.
AI model: Created by transfer learning from MobileNet-v2.
Physician A, B: Neurologist with over 10 years of clinical experience.
Physician C: Neurology resident.
^a Fisher's exact probability test is used to determine p values for accuracy between AI model and physician.

In this study, we developed AI models for evaluating DAT SPECT images using transfer learning of summed striatum images. Despite its simplicity and light load, the machine was able to interpret images with high accuracy, equivalent to that of experienced neurologists.

To diagnose PD accurately, it is necessary to determine that the presynaptic dopamine system is affected using functional imaging studies [15]. To evaluate the presynaptic dopamine system, ¹²³I-ioflupane SPECT is used widely in clinical practice as DAT SPECT and has become an indispensable tool for diagnosing PD [25]. While uptake of the isotope within the striatum is evaluated by visual assessment and an index SBR [26, 27], there is substantial inter-observer variability in the evaluation of visual morphology, making it difficult to detect subtle changes. Therefore, efforts have been made to apply AI to the interpretation of DAT SPECT images.

To create an entirely original CNN architecture for imaging analysis, specialized knowledge of machine learning, sufficient time, and a large amount of data are required. To address this challenge, a method called transfer learning is employed, in which a CNN is pre-trained on a large dataset such as ImageNet and then used as an initial model [28]. To overcome the lack of expertise in creating initial models and the scarcity of medical imaging data, we utilized transfer learning to develop an analysis system. As the pre-trained CNN architecture, we selected a lightweight model that can run on a personal computer in view of the ease of its implementation.

The accuracy of 3D analysis of DAT SPECT using AI has been reported to be as high as 96.0% [10]. Additionally, a report indicated that such methods can be used to differentiate PD, multiple system atrophy, and progressive supranuclear palsy [9]. These reports each describe the creation of new CNNs for 3D image analysis and demonstrate advanced image analysis techniques. However, the number of analyzed cases is very limited, and replication reports are scarce. This suggests the challenges and instability associated with 3D analysis of DAT SPECT images. In the current study, our AI model interpreted images with a similar accuracy as the 3D analysis models in the literature. Our analysis utilizes 2D images, which reduced the calculation load and enabled the operation of analysis with a commercially available computer.

A method involving extracting an image slice from DAT SPECT and performing image analysis has also been reported to achieve high accuracy [12]. This method is similar to that of present study in that it involves analyzing a single image. However, in our study, the processed image contained 3D information as it employed the summation of slices within a region of interest covering the striatum. Other studies have extracted multiple images, performed some form of processing, and analyzed the resulting 2D images [11, 13, 14], demonstrating highly accurate analytical capabilities, and their precision is comparable to that of 3D image analysis. The method we employed incorporated the concept of the SBR, which has been used in clinical practice, into the analysis images. Additionally, by learning not only the target region but also the entire image, we were able to obtain analytical results. There have been no previous reports of adding an entire image to the calculation range of the SBR. By adopting this method, we eliminated the process of specifying the analysis range, thereby simplifying and streamlining the analysis. This approach is novel, and has the potential to be applied as an analytical method in clinical practice in the future.

In this study, we utilized pre-trained CNN architectures such as AlexNet, GoogLeNet, ResNet-50, SqueezeNet, ShuffleNet, and MobileNet-v2 for transfer learning. These CNN architectures were selected based on reports of their usefulness in interpreting chest X-rays for pneumonia [4] and their availability in MATLAB. In addition, these are early models that marked the dawn of deep learning and their subsequent models, and they are characterized by their lightweight operation. Despite their light weight, all the models we utilized were capable of image analysis with high accuracy in the development cohort, and the same trend was observed in the validation cohort. In the data analysis results of the validation study, there were differences in accuracy among the AI models compared to the data analysis in the development study. Between the facilities, the DAT SPECT imaging conditions were almost identical. However, in the image creation process, attenuation correction was performed at our facility, but not at the collaborating research facility that generated the validation dataset. Attenuation correction reportedly yields data closer to the true image, but has a particular effect on contrast [29]. This may be a possible reason for the difference in accuracy, but both cases showed high accuracy. This observation shows that our AI model could perform accurate analysis regardless of whether adjustments have been made.

This study has several limitations. First, we used a small number of subjects, particularly in the Control group. In addition, while we summed DAT SPECT images to be analyzed and converted them into 2D images, converting 3D information into 2D may result in the loss of certain features. To address this issue, a method has been proposed that converts 3D images into 2D images without losing their 3D features [30]. It is desirable to explore the applicability of such methods and improve the analysis methods in the future. Furthermore, we did not verify the ability of our AI model to discriminate between PD and atypical parkinsonism.

In conclusion, the AI model developed in this study is simple, accurate, and user-friendly. Such a method may assist primary care physicians to screen patients to be referred to specialists.

CRediT authorship contribution statement

Atsunori Murao: Writing – original draft, Data curation, Formal analysis, Investigation, Methodology, Validation. Kazuhiro Hara: Writing - review & editing, Data curation, Funding acquisition, Investigation, Methodology. Shintaro Oyama: Writing - review & editing, Methodology. Aya Ogura: Writing - review & editing, Investigation. Yoshiyuki Kishimoto:Writing - review & editing, Investigation. Mai Hatanaka: Writing - review & editing, Investigation. Naotoshi Fujita: Writing - review & editing, Methodology. Misaki Sato: Writing - review & editing, Investigation. Ikuko Aiba: Writing - review & editing, Investigation. Katsuhiko Kato: Writing - review & editing, Methodology. Masahisa Katsuno: Writing - review & editing, Funding acquisition, Project administration, Supervision.

Disclosure statement

The authors declare no potential conflicts of interest.

Funding

This work was supported by Japan Society for the Promotion of Science (JSPS) KAKENHI Grant Number 24K10638, and MHLW Research on rare and intractable diseases Program Grant Number JPMH23FC1008.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Kalia LV, Lang AE. Parkinson's disease. Lancet. 2015;386(9996):896–912. 10.1016/S0140-6736(14)61393-3.
Dorsey ER, Bloem BR. The Parkinson Pandemic-A Call to Action. JAMA Neurol. 2018;75(1):9–10. 10.1001/jamaneurol.2017.3299.
Lundervold AS, Lundervold A. An overview of deep learning in medical imaging focusing on MRI. Z Med Phys. 2019;29(2):102–27. 10.1016/j.zemedi.2018.11.002.
Avola D, Bacciu A, Cinque L, Fagioli A, Marini MR, Taiello R. Study on transfer learning capabilities for pneumonia classification in chest-x-rays images. Comput Methods Programs Biomed. 2022;221:106833. 10.1016/j.cmpb.2022.106833.
Shao W, Rowe SP, Du Y. Artificial intelligence in single photon emission computed tomography (SPECT) imaging: a narrative review. Ann Transl Med. 2021;9(9):820. 10.21037/atm-20-5988.
Shiiba T, Arimura Y, Nagano M, Takahashi T, Takaki A. Improvement of classification performance of Parkinson's disease using shape features for machine learning on dopamine transporter single photon emission computed tomography. PLoS ONE. 2020;15(1):e0228289. 10.1371/journal.pone.0228289.
Nakajima K, Saito S, Chen Z, Komatsu J, Maruyama K, Shirasaki N, et al. Diagnosis of Parkinson syndrome and Lewy-body disease using ¹²³I-ioflupane images and a model with image features based on machine learning. Ann Nucl Med. 2022;36(8):765–76. 10.1007/s12149-022-01759-z.
Iwabuchi Y, Nakahara T, Kameyama M, Yamada Y, Hashimoto M, Matsusaka Y, et al. Impact of a combination of quantitative indices representing uptake intensity, shape, and asymmetry in DAT SPECT using machine learning: comparison of different volume of interest settings. EJNMMI Res. 2019;9(1):7. 10.1186/s13550-019-0477-x.
Zhao Y, Wu P, Wu J, Brendel M, Lu J, Ge J, et al. Decoding the dopamine transporter imaging for the differential diagnosis of parkinsonism using deep learning. Eur J Nucl Med Mol Imaging. 2022;49(8):2798–811. 10.1007/s00259-022-05804-x.
Choi H, Ha S, Im HJ, Paek SH, Lee DS. Refining diagnosis of Parkinson's disease with deep learning-based interpretation of dopamine transporter imaging. Neuroimage Clin. 2017;16:586–94. 10.1016/j.nicl.2017.09.010.
Jothi S, Anita S, Sivakumar S. Modified Exigent Features Block in JAN Net for Analysing SPECT Scan Images to Diagnose Early-Stage Parkinson's Disease. Curr Med Imaging. 2023 Jun;5. 10.2174/1573405620666230605092654.
Magesh PR, Myloth RD, Tom RJ. An Explainable Machine Learning Model for Early Detection of Parkinson's Disease using LIME on DaTSCAN Imagery. Comput Biol Med. 2020;126:104041. 10.1016/j.compbiomed.2020.104041.
Chen HY, Tsai YJ, Peng SJ. Interpretation of basal nuclei in brain dopamine transporter scans using a deep convolutional neural network. Nucl Med Commun. 2025;46(5):418–26. 10.1097/MNM.0000000000001963.
Buddenkotte T, Lange C, Klutmann S, Apostolova I, Buchert R. Fully automatic categorical analysis of striatal subregions in dopamine transporter SPECT using a convolutional neural network. Ann Nucl Med. 2025;39(6):618–30. 10.1007/s12149-025-02038-3.
Postuma RB, Berg D, Stern M, Poewe W, Olanow CW, Oertel W, et al. MDS clinical diagnostic criteria for Parkinson's disease. Mov Disord. 2015;30(12):1591–601. 10.1002/mds.26424.
Hattori M, Hiraga K, Satake Y, Tsuboi T, Tamakoshi D, Sato M, et al. Clinico-imaging features of subjects at risk of Lewy body disease in NaT-PROBE baseline analysis. NPJ Parkinsons Dis. 2023;9(1):67. 10.1038/s41531-023-00507-y.
Hiraga K, Hattori M, Satake Y, Tamakoshi D, Fukushima T, Uematsu T, et al. Plasma biomarkers of neurodegeneration in patients and high risk subjects with Lewy body disease. NPJ Parkinsons Dis. 2024;10(1):135. 10.1038/s41531-024-00745-8.
Tossici-Bolt L, Hoffmann SM, Kemp PM, Mehta RL, Fleming JS. Quantification of [123I]FP-CIT SPECT brain images: an accurate technique for measurement of the specific binding ratio. Eur J Nucl Med Mol Imaging. 2006;33(12):1491–9. 10.1007/s00259-006-0155-x.
Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Adv Neural Inf Process Syst. 2012;25:1097–105. doi.org/10.1145/3065386.
Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Going deeper with convolutions, 2015 IEEE Conference on Computer Vision and, Recognition P et al. (CVPR), Boston, MA, USA, (2015), pp. 1–9. doi.org/10.1109/CVPR.2015.7298594
He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition, Vision, Recognition P. (CVPR), Las Vegas, NV, USA, (2016), pp. 770–778. doi.org/10.1109/CVPR.2016.90
Iandola FN, Han S, Moskewicz MW, Ashraf K, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 mb model size, arXiv preprint arXiv:1602.07360 preprint (2016) 1–13.
Zhang X, Zhou X, Lin M, Sun J, ShuffleNet. an extremely efficient convolutional neural network for mobile devices, 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, (2018), pp. 6848–6856. doi.org/10.1109/CVPR.2018.00716
Sandler M, Howard A, Zhu M, Zhmoginov A, Chen L-C. MobileNetV2: inverted residuals and linear bottlenecks, 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, (2018), pp. 4510–4520. doi.org/10.1109/CVPR.2018.00474
Catafau AM, Tolosa E, DaTSCAN Clinically Uncertain Parkinsonian Syndromes Study Group. Impact of dopamine transporter SPECT using 123I-Ioflupane on diagnosis and management of patients with clinically uncertain Parkinsonian syndromes. Mov Disord. 2004;19(10):1175–82. 10.1002/mds.20112.
Yamamoto H, Arimura S, Nakanishi A, Shimo Y, Motoi Y, Ishiguro K, et al. Age-related effects and gender differences in Japanese healthy controls for [¹²³I] FP-CIT SPECT. Ann Nucl Med. 2017;31(5):407–12. 10.1007/s12149-017-1168-1.
Mizumura S, Nishikawa K, Murata A, Yoshimura K, Ishii N, Kokubo T, et al. Improvement in the measurement error of the specific binding ratio in dopamine transporter SPECT imaging due to exclusion of the cerebrospinal fluid fraction using the threshold of voxel RI count. Ann Nucl Med. 2018;32(4):288–96. 10.1007/s12149-018-1249-9.
Torrey L, Shavlik J. Transfer learning. Handbook of Research on Machine Learning Applications and Trends: Algorithms, Methods, and Techniques, IGI Global, 2010, pp. 242–264. 10.4018/978-1-60566-766-9.ch011
Fujita N, Abe S, Kato K, Aoki T, Sugimoto M. Investigation of Factors Relevant to the Differences among Facilities and SPECT Scanners for Standardization of Dopamine Transporter Scintigraphy (In Japanese), Kaku Igaku Gijutsu (Nuclear Medicine Technology). 36 (2016) 177–87.
Yoon H, Kang DY, Kim S. Enhancement and evaluation for deep learning-based classification of volumetric neuroimaging with 3D-to-2D knowledge distillation. Sci Rep. 2024;14(1):29611. 10.1038/s41598-024-80938-6.

Download PDF

Journal Publication

published 26 Nov, 2025

Read the published version in Annals of Nuclear Medicine →

Reviewers agreed at journal
12 Sep, 2025
Reviewers invited by journal
02 Sep, 2025
First submitted to journal
19 Aug, 2025
Editor assigned by journal
16 Aug, 2025

You are reading this latest preprint version

Examination of simple artificial intelligence-based analysis of dopamine transporter scintigraphy for supporting a diagnosis of Parkinson’s disease

Status:

Journal Publication

Version 1

Abstract

Introduction:

Methods

Results

Conclusion

Figures

Introduction

Methods

Study population

SPECT imaging

Image analysis

Creating a learning machine

Training algorithm for transfer learning

Use Deep Network Designer - MATLAB

-Classification layer

Training-progress settings

OutputNetwork -> last iteration

Confirmation of performance

Statistical analysis

Results

Characteristics of the case groups used in this study

Analysis of the test dataset using the AI models

Analysis of the PD validation group using the created AI models

Comparison of the most accurate AI model and clinicians

Discussion

Declarations

References

Status:

Journal Publication

Version 1