Objective: To develop a self-supervised deep learning model using ¹⁸F-FP-CIT PET images from patients with Parkinsonism and to evaluate the model’s performance in the differential diagnosis of Parkinsonism.

Background: Dopamine transporter imaging is key in distinguishing Parkinsonism types, complementing diagnoses from patient histories and exams. Self-supervised learning models, learning good representation of images without labels, have garnered interest. Yet, studies on these models using dopamine transporter images from Parkinsonism patients are scarce.

Method: ¹⁸F-FP-CIT PET images of patients with Parkinsonism were collected from January 2005 to March 2022. These images were used to train our self-supervised, deep generative model, the Hierarchical Wavelet Diffusion Autoencoder (HWDAE), which combines wavelet transformation and the Hierarchical Diffusion Autoencoder [1]. The pretrained semantic encoder of HWDAE was further trained using either fine-tuning or linear probing method to address a three-class classification among Parkinson’s disease (PD), multiple system atrophy (MSA), and progressive supranuclear palsy (PSP). The model’s performance was evaluated using 5-fold cross validation and compared with two self-supervised (Denoising Diffusion Autoencoders (DDAE) [2], SimMIM [3]) and one supervised model (3D Resnet34).

Results: We utilized 2672 PET images to pretrain the self-supervised models for classification. For the classification, 231 training images (127 PD, 118 MSA, 44 PSP), 72 for internal validation (32 PD, 29 MSA, 11 PSP), and 291 for external validation (261 PD, 22 MSA, 8 PSP) were included. The external validation set consisted of images from the same medical center but obtained with a different PET/CT scanner. HWDAE trained with linear probing achieved an average AUC of 0.83 ± 0.05 (internal) and 0.88 ± 0.03 (external), while fine-tuning showed improvement with 0.92 ± 0.02 (internal) and 0.92 ± 0.03 (external). In comparison, DDAE scored 0.91 ± 0.03 (internal) and 0.89 ± 0.04 (external), SimMIM had lower performance with 0.74 ± 0.04 (internal) and 0.69 ± 0.08 (external), and the 3D Resnet34 model achieved consistent results of 0.91 ± 0.02 (internal) and 0.91 ± 0.04 (external).

Conclusion: We developed a self-supervised model with dopamine transporter imaging and showed its model superiority in a classification task of patients with Parkinsonism.

References: 1. Lu Z, Wu C, Chen X, et al. Hierarchical diffusion autoencoders and disentangled image manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024:5374-5383.
2. Xiang W, Yang H, Huang D, Wang Y. Denoising diffusion autoencoders are unified self-supervised learners. Proceedings of the IEEE/CVF International Conference on Computer Vision; 2023. p. 15802-15812.
3. Xie Z, Zhang Z, Cao Y, et al. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022:9653-9663.

« Back to 2024 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/differential-diagnosis-of-parkinsonism-with-18f-fp-cit-pet-using-a-self-supervised-deep-learning-model/