Related Experiment Videos
Hierarchical deep learning-radiomics fusion for differentiating Parkinson's disease from atypical Parkinsonism using
Jun Kim1, Young Jae Kim2, Kawngwoo Park3
1Department of MetaBioHealth, School of Medicine, Sungkyunkwan University, Suwon, Republic of Korea; Department of Biomedical Engineering, College of IT Convergence, Gachon University, Gil Medical Center, Seongnam-si, Gyeonggi-do, Republic of Korea.
Background:
Parkinson's disease (PD) and atypical Parkinsonism (AP) are challenging to differentiate, particularly in the early stages, because of overlapping motor symptoms and similar dopaminergic neurodegenerative patterns. We developed an interpretable artificial intelligence (AI) framework integrating deep learning and radiomics using automatically extracted striatal regions of interest (ROIs) from FP-CIT imaging for the differential diagnosis of Parkinsonian disorders.
Methods:
A total of 228 participants, including healthy controls (HC), patients with PD, and patients with AP, were retrospectively analyzed. Striatal ROIs were automatically extracted from FP-CIT images using a hotspot-based bounding-box algorithm. Radiomics features and deep learning-derived imaging features were integrated using a hierarchical late fusion framework. Hierarchical classification was performed in two stages: HC versus Parkinsonism, followed by PD versus AP. Model interpretability was evaluated using Grad-CAM and SHAP analyses.
Results:
Across five random seeds, the validation-selected late-fusion model achieved a mean held-out test accuracy, F1-score, and ROC-AUC of 0.906, 0.942, and 0.969, respectively, for Stage 1 (HC vs. parkinsonism) and 0.815, 0.669, and 0.880 for Stage 2 (PD vs. AP). The ablation analysis showed modest improvements in mean accuracy and F1-score over the deep learning-only model in both stages. Grad-CAM provided qualitative visualizations of the image regions contributing to model predictions, while SHAP identified stage-dependent radiomics importance profiles primarily involving voxel-intensity distributions and texture heterogeneity.
Conclusion:
The proposed framework demonstrated promising internal performance for the hierarchical classification of Parkinsonian disorders and provided exploratory information regarding the image regions and radiomics features contributing to model predictions. These findings support the feasibility of integrating deep learning and radiomics for FP-CIT image analysis; however, independent multicenter validation is required before its external generalizability and clinical utility can be established.