Gabriel Jiménez, PhD, of the Paris Brain Institute, presented their research that leverages whole slide imaging (WSI) and graph neural networks to analyze the spatial organization of Alzheimer’s disease pathology.[1] The study identified topographical patterns of tau aggregates that distinguish between classic and rapid progressive forms of the disease, offering novel insights into disease progression and heterogeneity.

“The clinical manifestations of Alzheimer’s patients are quite diverse, which presents a significant challenge for us,” Jiménez explained during his presentation. “Current in vivo imaging technologies cannot match the resolution of microscopic images, limiting our understanding of the disease.”

The research team developed a pipeline that converts WSIs into graph representations of tau aggregate distributions, allowing for analysis of both local features and global spatial patterns of pathological deposits across different cortical layers. The pipeline began with an image segmentation step, during which pathologist insights are crucial.

“We started with the annotation refinement tool because the first task we need to do is segmentation,” Jiménez noted. “We need to locate the tissue, and then we propose that pathologists and other experts should be included in the loop so they can refine their rotation.”

Jiménez discussed the several technical challenges the team encountered during the development of their pipeline. Manual annotation variability among pathologists reached approximately 30%, and different imaging parameters affected the results, with antibody selection and scanner variations influencing the model performance. To address these challenges, the researchers implemented a graph-based framework that could integrate information across entire tissue sections. The approach incorporated both the morphological properties of individual plaques and the topological features of their spatial relationships, allowing the team to identify distinct distribution patterns of pathology in the six cortical layers in the frontal cortex between rapid and classic Alzheimer’s disease.

“When we use the embeddings, we see that they follow the distribution of the layers of the brain,” Jiménez said. “With the embeddings and the random forest classification, we found that different layers yield different results for the stratification of patients.”

Rapid progressive Alzheimer’s disease showed denser networks predominantly affecting the middle cortical layers, particularly layer 3, whereas classic Alzheimer’s disease demonstrated a more dispersed pattern involving both superficial and deep layers.

The research team is currently working to enhance the model by incorporating additional data types. “Something that we are working on now is integrating additional information in the nodes because currently, we are only integrating the position of the plaques in the tissue,” Jiménez said. “We can also integrate the morphological properties of the plaques and the relationship between the plaques and tangles in the tissue.”

Future directions include extending the methodology to other neurodegenerative conditions, such as frontotemporal dementia and Parkinson’s disease. The team also hopes to correlate their findings with in vivo imaging data, though Jiménez acknowledged that obtaining matched datasets is challenging.Gabriel Jiménez, PhD, research engineer at Paris Brain Institute, France.

[1] Gabriel Jiménez, Unravelling the topographical organization of brain lesions in variants of Alzheimer’s disease progression. Presented at SPIE 2025 Digital and Computational Pathology conference, February 19, 2025; San Diego, CA.