Enhancing interpretability of AI with radiomics-based deep neural network: proof of concept in the classification of Parkinsonian syndromes with 18F-FDG PET imaging

Le résumé fourni par la source

Objective: Interpretability and reproducibility remain major challenges in applying deep neural network (DNN) to neuroimaging-based diagnosis. This study proposes a radiomics-guided dual-channel deep neural network (RDDNN) to improve feature transparency and enhance clinical understanding in the classification of Parkinsonian syndromes. Methods: In this bi-centric study, we analysed two independent cohorts comprising 1,275 patients with idiopathic Parkinson's disease (IPD), multiple system atrophy (MSA), and progressive supranuclear palsy (PSP), alongside 223 healthy controls from Huashan Hospital and 90 patients with IPD, MSA, and PSP (34IPD, 17MSA, 39PSP) from the University Hospital Munich. It is a re-analysis of well-studied Chinese and German cohorts of 18F-fluorodeoxyglucose Positron emission tomography (FDG-PET) imaging of parkinsonian patients and the FDG scans were of 10-min static acquisition at 60 min post FDG injection and normalized against whole brain activity. The RDDNN model combines local features extracted via dilated convolutional networks and global features derived from Transformer-based self-attention networks. Model performance was evaluated using classification metrics and compared to radiomics and DNN approaches. The model's outputs were also compared with nuclear medicine specialists' visual assessments to assess interpretability and time efficiency. Furthermore, SHapley Additive Explanations (SHAP), Layer-wise Class Activation Mapping (Layer-CAM), and Rollout Attention Map (RAM) were employed to evaluate which features played the most critical roles in the model's final classification decisions after supervised training, and to examine how both networks spatially corresponded to known brain connectivity regions. Results: In the internal blind-test cohort, the RDDNN achieved high accuracy (AUC = 0.99, accuracy = 0.98). SHAP and correlation analyses jointly indicated complementary information across channels, some of which were clinically interpretable. In the external cohort, the model maintained robust performance (AUC = 0.94, accuracy = 0.81), with consistent feature patterns across populations. The model significantly reduced evaluation time compared to nuclear medicine specialists' readings (p < 0.001), and the heatmaps showed disease-specific activation in anatomically relevant regions for IPD, MSA, and PSP. Conclusion: The RDDNN framework provides a clinically interpretable and reproducible DNN solution for classifying Parkinsonian disorders. By integrating radiomics and attention-based modeling, it enhances lesion localization, supports clinical decision-making, and offers performance comparable to human specialists-while substantially improving diagnostic efficiency.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.