DepthParkNet: A 3D Convolutional Neural Network with Depth-Aware Coordinate Attention for PET-Based Parkinson's Disease Diagnosis

Positron Emission Tomography (PET) imaging is essential for accurately diagnosing Parkinson's Disease (PD). With the advancement of deep convolutional networks, 3D convolutional neural networks (CNNs) have been extensively utilized to analyze PET images. However, the limited availability of PET imaging data represents a significant bottleneck in this field. Training on small datasets could restrict the performance of 3D CNNs. Moreover, models trained on small datasets tend to exhibit weak generalizability, making it challenging to handle variations in PET images caused by clinical factors. This paper proposes depth-aware coordinate attention that integrates clinical prior knowledge. It's composed of two modules: the coordinate attention module and the depth-aware attention module. The coordinate attention module could preserve positional information while capturing long-range dependencies. Meanwhile, the depth-aware attention module aims to reveal the relationships among various depths within a PET image. These innovations enable our DepthParkNet to extract critical features from limited PET imaging data, thus enhancing its performance. Additionally, this paper proposes an augmentation pipeline named PDaug, which includes three transformations aimed at improving the generalizability of the model. The proposed model is validated on two datasets, demonstrating outstanding performance with balanced accuracy of 99.17% and 98.13%.