Deep Learning-Based Classification of Temporal Stages of AT8-Labeled Tau Pathology After Experimental Traumatic Brain...

Area of Science:

• Neuropathology
• Computational Neuroscience
• Biomedical Imaging

Background:

• Tauopathies are neurodegenerative diseases characterized by abnormal tau protein accumulation.
• Quantifying early and intermediate tauopathy stages is difficult due to subtle morphological changes.
• Traumatic brain injury (TBI) can induce tauopathy, but its temporal progression is underexplored.

Purpose of the Study:

• To evaluate a deep learning framework for classifying temporal stages of tauopathy progression.
• To assess the performance of different convolutional neural network (CNN) architectures in this task.
• To apply this framework to AT8-stained cortical micrographs from a controlled TBI mouse model.

Main Methods:

• Three CNN architectures (custom, InceptionV3, DenseNet) were trained on AT8-stained micrographs.
• Images were categorized into four post-TBI stages: 1 day, 1 week, 1 month, and 3 months.
• Data preprocessing included normalization, augmentation, and oversampling, with performance evaluated using cross-validation.

Main Results:

• DenseNet achieved the highest accuracy (70.9%) and macro-F1 score (0.68), with excellent discrimination for the 1-week stage (F1=0.95).
• All models struggled with the earliest stage (1 day) and showed partial overlap in later stages (1-3 months).
• Results suggest deep learning can automate tauopathy staging in preclinical histology.

Conclusions:

• Deep learning, especially transfer learning models like DenseNet, offers a scalable method for automated temporal staging of tauopathy.
• The framework provides a foundation for digital neuropathology but requires larger datasets for generalization.
• Future work should focus on multi-center data and slide-level modeling for improved early detection and treatment evaluation.