Decoding the dopamine transporter imaging for the differential diagnosis of parkinsonism using deep learning

The accurately and timely differential diagnosis of parkinsonian disorders remains challenging due to overlapping symptoms, especially in the early stage, between patients with idiopathic Parkinson’s disease (IPD) and atypical parkinsonian syndromes (APS), e.g., multiple system atrophy (MSA) and progressive supranuclear palsy (PSP) [1]. Pathological examination results show that approximately 20–30% of patients with MSA or PSP are initially misdiagnosed as IPD in clinical practice [1]. Therefore, developing an accurate computer-aided diagnosis method for differential diagnosis of parkinsonian disorders is of great value to avoid unnecessary testing and inappropriate medicines and thus leads to better therapeutic strategies.

The dopamine transporter (DAT) imaging such as [¹¹C]CFT positron emission tomography (PET) and [¹²³I]FP-CIT single-photon emission computed tomography (SPECT) (DaTscan) can reflect the subject’s dopaminergic degeneration and therefore is a powerful diagnostic tool [2]. Nowadays, striatal DAT quantification together with visual analysis is utilized as a standard practice in clinical studies [3]. However, according to current knowledge, DAT imaging has not been confirmed to be suitable for the reliable differentiation of IPD and APS subtypes based on conventional quantitative analyses such as the DAT binding ratio (BR) quantification [4‐6]. These conventional quantitative analyses normally focus on specific brain regions including putamen and caudate, and each region is represented by its mean counts, hence underutilizing the global information of entire DAT scans, especially the distribution of the uptakes within each region as well as correlations of different brain regions.

The deep neural network has been demonstrated to be able to decode in-depth features automatically and effectively from the data [7‐11], which has the potential to discover more comprehensive information and update its parameters specifically for the differential diagnosis of parkinsonism. Moreover, deep learning may assist the conventional quantitative analysis or radiomics analysis [12‐14]. The potential of deep learning has been revealed in the analysis of DAT imaging [15‐18]. Choi et al. introduced the deep learning method to refine the imaging diagnosis of Parkinson's disease based on the FP-CIT SPECT scans[16]. Wenzel et al. reported that the deep neural network can be trained to be robust to variable image characteristics for the classification of the FP-CIT SPECT [17]. Currently, these works mainly focused on the differentiation between PD and healthy controls but did not evaluate the potential of deep learning to solve a more challenging task, i.e., the differential diagnosis of IPD from APS. Utilizing unsupervised dimension reduction method and hierarchical clustering, Suh et al. divided FP-CIT PET scans into multiple groups and then evaluated the correlation between certain clusters and specific Parkinson symptoms, which depicted the heterogenous dopaminergic neurodegeneration patterns in parkinsonian [19]. However, it cannot directly characterize the probabilities of each parkinsonian syndrome and provide a diagnosis prediction when given an unseen scan.

In this study, we leveraged deep learning to extract informative imaging signatures from [¹¹C]CFT PET scans to support the differential diagnosis of parkinsonian syndromes. A 3D deep residual convolutional neural network (termed as DAT-Net) was proposed, which can get access to the entire DAT image and involve the uptake distribution, content and context information among different regions. A large multi-cohort dataset of DAT imaging was collected to develop the DAT-Net and then to evaluate its performance. Furthermore, we investigated the decision mechanism of the deep learning network based on the state-of-the-art full-gradient saliency map method [20], which provides a view to understand the deep neural network and reveal the functional abnormal regions of patients with different syndromes indicated on the [¹¹C]CFT PET scans.

Evaluated on one of the largest available datasets with 1017 subjects, our preliminary results demonstrated that the potential of the [¹¹C]CFT PET scans for differentiating IPD and APS subgroups based on the proposed DAT-Net, which might benefit from deep learning’s ability for decoding critical information from DAT imaging. In current clinical practice, DAT imaging is considered unsuitable for the reliable differentiation of IPD and APS subtypes such as MSA and PSP[29, 30]. While the conventional differential diagnosis referring to DAT imaging is based on quantitative analysis such as the DAT binding ratio (BR) quantification [4‐6] in specific regions such as putamen and caudate, the numerous information of DAT imaging was neglected. In the current study, taking advantage of deep-learning method, we were able to dig deeper into this classical functional imaging modality and successfully expanded its significance for disease diagnosis with DAT-Net, DL-BR and DL-radiomics.

The performance of the DAT-Net was evaluated both in the cross-validation stage and blind-test stages. The success of the neural network mainly benefited from its capacity to access the global information of entire PET scans and analyze multiple regions as well as their correlation simultaneously, which was different from the traditional method that only focused on certain slices and certain regions. The relatively comparable performance in these two stages showed the robustness of our proposed network between different cohorts. The high accuracy on the patients with short symptom duration (Fig. 3, Supplementary Table 2) and patients at baseline (Fig. 4, Supplementary Table 3) suggested the potential of DAT-Net for early diagnosis. Patients with longer symptom durations were supposed to have more extensive changes in the brain as disease progression. Our network obtained comparable performance between patients with short symptom duration/at baseline and with long symptom duration/at follow-up, which also implied that the proposed network was sensitive to brain changes on the DAT imaging scans, i.e., even slight changes in the early stage can be recognized by the DAT-Net to provide a similar accurate diagnosis, compared to that made after referring to more significant changes at follow-up (longer symptom duration). Besides, our proposed network achieved remarkable performance in the differential diagnosis of parkinsonian patients (with IPD or APS) from NC, which confirmed the ability of DAT imaging for the diagnosis of parkinsonian patients from NC due to the significant striatal DAT loss in the image.

The saliency maps suggested that the network paid attention mainly to putamen, caudate and midbrain, which meant that these regions were assigned higher importance scores and therefore contributed mainly to the final prediction of the network, though the remaining regions also showed contribution. The putamen and caudate, which were regarded as the leading corresponding brain regions for disease progression [31, 32], were the most common included regions in the traditional conventional quantitative analysis [33, 34]. For the midbrain, as a vital structure in dopamine signaling, the heterogeneity of dopamine neurons was considered to underpin the variety of clinical symptoms [35]. Moreover, [¹¹C]CFT also displays a high affinity for serotonin transporters (SERT) in addition to DAT in the midbrain and previous studies have shown that midbrain SERT distribution is significantly different between PD and MSA-P groups or between PD and PSP groups [34, 36], which may be another important factor suggesting the DAT-Net to pay attention to the tracer binding in midbrain for the differential diagnosis. All these previous findings supported that the detected regions in the saliency maps were in accordance with the key structures of the underlying pathological mechanisms.

Comparing the performance of the conventional BR and our designed DL-BR among IPD, MSA and PSP groups, we found the conventional putamen and caudate BR had limited potential for differential diagnosis, which was consistent with existing research results [29, 30]. However, there were significant differences among IPD, MSA and PSP groups if evaluating with the new designed DL-BR. This result was in line with our previous DAT-Net analysis that the neural network can access the entire PET scan and then learn the specific regions with distinguishable PET information for the differential diagnosis of IPD, MSA and PSP. Similarly, compared to conventional radiomics features, more DL-radiomics features showed significant differences between different groups (IPD vs MSA, IPD vs PSP and MSA vs PSP), which also indicated the advantage of the learned specific regions by the DAT-Net. Subregional patterns analysis of dopamine transporter loss was suggested as a potential way to improve the differential diagnosis of parkinsonism [37]. The improved performance of the subregions defined by deep learning in this study may provide a complementary tool to identify more efficient subregional patterns. Furthermore, the neural network can not only locate the diagnosis-informative regions but also assign weights on each voxel within these regions. Therefore, it may be more supportive than DL-BR and DL-radiomics in the differential diagnosis. To be specific, compared to conventional quantitative analysis which assigned the same weight on the uptake of each voxel when analyzing certain regions, the neural network allowed assigning different automatically learnt optimal weights on different voxels (Fig. 5). These weights may reflect the inter-correlation among regions, which may be dependent on the pathogenesis of different types of parkinsonism. While all suffered from dopaminergic dysfunction, previous studies showed that IPD, MSA and PSP had different preferential subregional decreases in striatal DAT binding[37] and different speeds of dopaminergic degeneration [38]. Considering the different directions and different speeds during the progression of dopamine transporter loss, future investigation of these interrelations may assist the understanding of the pathway behind the disease. Overall, the combination of the diagnosis-informative sub-region and the interrelation-determined weights has improved the ability of the network in the differential diagnosis of IPD, MSA and PSP.

Unsurprisingly, our experiments also illustrated that leveraging multi-modality data slightly outperformed only utilizing image modality, which might be due to that the image-only modality itself already achieved a relative high accuracy for differentiation. While the most important three features were all extracted from the DAT imaging scans, we inferred that the image modality accounted for the main contribution of the proposed multi-modality method and the demographic and clinical features were beneficial for improving the diagnosis performance and robustness.

As a data-driven method, data played a central role in the development of neural networks. Sufficient training data were helpful for the network to learn features of diversity cases and therefore could improve the performance of the network and prevent them from over-fitting. Although the relatively large dataset in this study allowed for a comprehensive understanding of this neural network within IPD and APS, the influence of the physical complexity of imaging data on such models remained to be further explored. And therefore, we would like to evaluate the performance of our DAT-Net when addressing scans obtained from different devices. Another limitation of the present study was that the entire experiment was based on a retrospective cohort and multi-center prospective studies were still needed to further confirm the protentional of this method in the differential diagnosis of parkinsonism based on DAT imaging. Furthermore, we utilized one possible multi-modality fusion method to synthesize multimodal information in this work. In the future, other fusion methods such as gating-based attention-based late fusion [39] will be leveraged to further improve the multimodal diagnosis performance. Another interesting future work is to evaluate the potential of utilizing the pre-trained DAT-Net (trained on [¹¹C]CFT PET) for the differential diagnosis of parkinsonism on FP-CIT SPECT. There exist large gaps of tracer, modality and ethnics between the two modalities. The domain-adversarial training strategy [40] may have the potential to alleviate the influence of the modality gap and further extend the proposed method of this study.

In this paper, we developed a 3D deep residual convolutional neural network for the automated differential diagnosis of parkinsonism based on the [¹¹C]CFT PET scans. The experiment results in the cross-validation phase and the blind test phase demonstrated the relatively high accuracy and the generalization ability of the proposed method. The saliency maps of the deep neural network, which indicated the functional basis and distinguishable regions recognized by the network, were consistent with known neuropathological processes in parkinsonism. These findings suggest that employing the deep neural network in the analysis of DAT scans has the potential to assistant neurologists in making reasonable diagnoses of the parkinsonian syndromes at an early stage.