AG-PDCnet

AG-PDCnet: An Attention Guided Parkinson’s Disease Classification Network with MRI, DTI and Clinical Assessment Data

Sushanta Kumar Sahu, Ananda S. Chowdhury

Department of Electronics and Telecommunication Engineering, Jadavpur University, Kolkata, India.

Acepted in CVIP 2023

Parkinson’s disease (PD) is the second most neurodegenerative disorder, which is prevalent worldwide. In this paper, we propose AG-PDCNet, an Attention Guided multi-class multi-modal PD Classification framework. In particular, we combine clinical assessments with the Neuroimaging data, namely, MRI and DTI. The three classes considered for this problem are PD, Healthy Controls (HC) and Scans Without Evidence of Dopamine Deficiency (SWEDD). Four CNNs, each boosted with an attention mechanism, are trained on gray matter (GM) and white matter (WM) from the MRI, and mean diffusivity (MD) and fractional anisotropy (FA) from the DTI. XGboost is employed for classification from the clinical data. At the decision level, the outputs of all the five models, four CNNs and the XGboost, are fused with an optimal weighted average fusion (OWAF) technique. Publicly available PPMI database is used for evaluation, yielding an accuracy of 96.93 % for the three-class classification. Extensive comparisons, including ablation studies, are conducted to validate the effectiveness of our proposed solution.

Our main contributions are listed below:

We addresses the challenge of classifying PD, SWEDD and HC subjects. The inclusion of the SWEDD class significantly increases the complexity of the classification task.
Our framework integrates MRI, DTI and clinical assessments data to improve classification accuracy. We fuse the outputs of ML and DL based models trained on neuroimaging and clinical assessments data using late fusion.
We incorporate an attention mechanism to selectively focus on informative features or regions within the data. This enhances the discriminative power of the classification model by capturing subtle patterns and correlations necessary for accurate PD classification.

CODE