Soft Attention Based DenseNet Model for Parkinson's Disease Classification Using SPECT Images.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning aids in diagnosing neurological disorders like Parkinson's disease (PD).
• DaTscan (SPECT imaging) is crucial for PD detection.
• Identifying specific brain regions is key for accurate PD diagnosis.

Purpose of the Study:

• Develop a convolutional neural network (CNN) for PD diagnosis using DaTscan.
• Enhance PD detection by identifying regions of interest (ROIs) through feature extraction.
• Improve diagnostic accuracy for Parkinson's disease.

Main Methods:

• Utilized DenseNet-121 architecture with a soft-attention block for classification.
• Analyzed 1,390 DaTscan images from PD patients and healthy controls.
• Employed Soft Attention Maps and feature maps for visual ROI analysis.

Main Results:

• Achieved 99.2% accuracy, 99% AUC-ROC, 99.2% sensitivity, 99.4% specificity, and 99.1% F1-score.
• Exceeded previous research benchmarks in diagnostic performance.
• Soft-attention maps highlighted putamen and caudate regions, crucial for PD identification.

Conclusions:

• Deep learning effectively diagnoses Parkinson's disease using DaTscan images.
• The model accurately identifies putamen and caudate areas, distinguishing PD from normal cohorts.
• High accuracy and sensitivity demonstrate the potential of this AI framework in clinical settings.