Skip to content

AI Models to Assess Histopathologic Markers of Alzheimer's Disease, Cerebral Amyloid Angiopathy

Key findings

  • Massachusetts General Hospital researchers trained and validated six customizable convolutional neural networks (CNNs) to analyze various histopathological markers common in both Alzheimer's disease and cerebral amyloid angiopathy (CAA)
  • Precision, sensitivity and F1 score were good (>80%) to excellent (>90%) for all CNNs compared with analyses by expert human raters
  • Results derived from the models trained to detect amyloid-ß or iron were consistent with previously acquired semiquantitative scores in the same dataset
  • The CNNs produce continuous variables that enable application of more complex statistical models; in a linear mixed effects model the researchers reproduced the previously reported association between leptomeningeal CAA and iron-positive cells
  • The CNNs were built with the aid of an online platform that does not require the user to have knowledge of artificial intelligence or programming

Traditionally, amyloid-β (Aβ) accumulation and other neuropathological markers of Alzheimer's disease (AD) and cerebral amyloid angiopathy (CAA) are quantified using visual semiquantitative scores or a manual count. However, these scores are subjective and time-consuming to standardize across centers, and they cannot be used in complex statistical models of disease pathophysiology.

To overcome these deficiencies, Valentina Perosa, MD, research fellow, and Susanne J. van Veluw, PhD, assistant professor in the Department of Neurology at Massachusetts General Hospital, and colleagues developed artificial intelligence–based models that identify characteristic markers of AD and CAA. They describe their workflow and results in Acta Neuropathologica Communications.

Study Methods

Convolutional neural networks (CNNs), a form of deep learning, recognize patterns from medical images or other visual data. Using digitized histopathological slides from brains of patients with AD and CAA, the team trained six models to detect different features of interest:

  • Aβ model—Percentage of leptomeningeal CAA area, cortical and white matter CAA area, and cortical and white matter Aβ plaque area
  • Iron model—Cortical density of iron-positive cells (indicative of hemorrhage and siderosis)
  • Fibrin model—Percentage area of fibrin-positive vascular tissue and density of fibrin-positive cells in cortex and white matter (indicative of blood–brain barrier leakage)
  • GFAP model—Cortical density of reactive astrocytes (indicative of neuroinflammation)
  • CD68 model—Density of activated microglia in cortex and white matter (indicative of neuroinflammation)
  • Calcium model—Percentage of calcium-positive tissue area, vascular tissue area, extracellular area and cellular area in cortex and white matter (have been observed in the hippocampus of AD patients and those with severe CAA)

Model Performance

The CNNs were validated on slides different from those on which the models had been trained. Compared with analysis by three expert human raters, the precision, specificity and F1 score were very good (>80%) to excellent (>90%) for all models. The F1 score is the harmonic mean of precision and sensitivity.

Semiquantitative Scores and CNN-derived Measures

Results derived from the Aβ and iron models were consistent with previously acquired semiquantitative scores in the same dataset (P<0.001 for all comparisons):

  • Scores of cortical and leptomeningeal CAA strongly correlated with CNN-derived cortical (P=0.82) and leptomeningeal (P=0.75) CAA percentage area
  • Scores of cortical Aβ plaque strongly correlated with CNN-derived Aβ plaque percentage area (P=0.84)
  • Scores of iron deposits strongly correlated with CNN-identified density of iron-positive cells (P=0.72)

Clinical CAA Cohort

Continuous measures obtained by the CNNs permit construction of more complex statistical models that assess the interplay of pathological markers.

For example, in a linear mixed-effects model that included data on 13 CAA patients, the researchers reproduced the association between leptomeningeal CAA and iron-positive cells, which was reported in Brain, while controlling for cortical CAA as a potential other contributor.

No Technical Knowledge Required

CNNs have previously been trained to recognize and quantify CAA and Aβ plaques. The important advantage of the new CNNs is that they can concurrently assess additional markers in the same dataset to assess complex disease associations.

The team used an online platform, Aiforia, that supports the development of CNNs for histopathological analysis without requiring the user to understand artificial intelligence or programming.

The CNNs could be customized to incorporate other neuropathological markers. Standardized assessment of markers across research centers would allow pathophysiological questions in neurodegenerative disease to be addressed in a harmonized way and on a larger scale.

Learn more about the Department of Neurology

Learn more about the Van Veluw Lab for Neuroimaging


In a study with major implications for drug development in Alzheimer's disease, Raja Bhattacharyya, PhD, and Rudolph E. Tanzi, PhD, of the McCance Center for Brain Health, and colleagues have determined how neurotoxic amyloid-β is generated in axons and how its production can be reduced.


Ibai Diez, PhD, and Jorge Sepulcre, PhD, DMSc, MD, of the Gordon Center for Medical Imaging, explain how neuroimaging findings are being combined with genetic data—research that might lead to disease-modifying therapies for neurodegenerative diseases.