Prediction of cognitive impairment in morphologically- and immunohistochemically-stained whole slide images reveals white matter alterations
Date:
Background: Age-related cognitive impairment is complex and multifactorial, with numerous underlying and frequently co-morbid pathological correlates. Given that conventional measures are imperfect predictors of cognitive impairment, improved approaches are urgently needed to identify correlated histopathological changes in an unbiased way.
Methods: We used a weakly supervised deep learning algorithm, clustering-constrained-attention multiple-instance learning (CLAM), on digitized whole slide images of human brain tissue sections from a group of elderly individuals without significant amyloid-beta deposition to predict the presence or absence of cognitive impairment. The brain tissue sections were from the hippocampus and frontal cortex and were either morphologically stained with LH&E or immunohistochemically stained for hyper-phosphorylated tau. We used attention analysis to pinpoint the underlying architectural and cellular features that the top-performing models used for the predictions.
Results: Despite noisy labels of cognition, our trained models were able to predict the presence of cognitive impairment with an accuracy that significantly beat chance, in both the morphologically stained and immunohistochemically stained whole slide images. Attention-based analysis of the features most associated with cognitive impairment in both types of whole slide images analyzed suggests that the models identified white matter alterations in donors with cognitive impairment.
Conclusions: The deep learning method for the prediction of cognitive impairment that we present, which appears to be minimally influenced by the human-identified visually salient aspects of whole slide images, may help to discover unexpected aspects of the pathophysiology associated with cognitive impairment in the elderly.