Diagnosis of Parkinsonian Syndromes in 18F-FDG PET Images Based on 3D CNN-LSTM-Attention

Parkinson's syndrome (PS) is a group of neurodegenerative disorders characterized by motor impairments, including idiopathic Parkinson's disease (IPD), progressive supranuclear palsy (PSP), and multiple system atrophy (MSA). Accurate differentiation between these subtypes is crucial for optimizing treatment due to significant differences in their pathological mechanisms and therapeutic responses. 18F-FDG PET, a functional imaging technique, provides reliable early-stage detection for neurodegenerative diseases. This study aims to develop a deep learning model that automatically classifies PS subtypes from 18F-FDG PET images and generates corresponding diagnostic reports in Chinese. The dataset from Fudan University-affiliated Huashan Hospital's Parkinson's PET Imaging Database (HPPI, China) includes 1035 18F-FDG PET images (574 IPD, 215 PSP, 246 MSA) and 543 Chinese diagnostic reports (265 IPD, 127 PSP, 151 MSA). The model uses a 3D convolutional neural network (3D CNN) for image feature extraction, combined with long short-term memory (LSTM) and attention mechanisms for report generation. Experimental results show that the model achieved an accuracy of 0.943 and a recall of 0.941 in the three-class classification task. For structured text generation, BLEU, ROUGE-1, ROUGE-2, ROUGE-L, and METEOR scores were 0.684, 0.934, 0.883, 0.934, and 0.661, respectively. Additionally, we evaluated the model's interpretability using activation maps (Grad-CAM., Grad-CAM++., and Layer-CAM) and demonstrated that Layer-CAM outperformed in both performance and stability, offering enhanced model interpretability and clinical decision support.