Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session E

Poster R11 - Neuropathological Whole Slide Image Analysis Using Attention-Based Weakly-supervised Learning

Saturday, October 26, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Shuying Li, PhD (she/her/hers)

Postdoctoral Associate
Boston University
Boston, Massachusetts, United States

Co-Author(s)

JC

Jonathan Cherry, PhD

Assistant Professor
Boston University School of Medicine, United States
MM

Maxwell Malamut, MS

Master's student
Boston University, United States
AM

Ann McKee

Professor
Boston University School of Medicine, United States

Primary Investigator(s)

Lei Tian, PhD

Associate Professor
Boston University, United States

Introduction: Our dataset included whole slide images (WSIs) from subjects categorized into four groups: normal controls without repetitive head injury (NC), patients exposed to RHI without developing CTE (RHI), patients with early-stage CTE (Low CTE), and patients with advanced-stage CTE (High CTE). We analyzed WSIs from the frontal cortex region using two stains: pathological AT8 immunostaining and structural Luxol fast blue, hematoxylin, & eosin (LH&E) staining. Attention-based multiple instance learning models were trained to detect subtle changes across CTE stages: NC to RHI, RHI to Low CTE, and Low CTE to High CTE. This weakly-supervised approach eliminates the need for laborious manual annotations by classifying WSIs based on slide-level diagnoses rather than pixel-level annotations. The self-learned attention scores were used to plot attention maps that highlight critical regions in the slides, facilitating the identification of relevant pathological features.

Materials and

Methods: Our dataset included whole slide images (WSIs) from subjects categorized into four groups: normal controls without repetitive head injury (NC), patients exposed to RHI without developing CTE (RHI), patients with early-stage CTE (Low CTE), and patients with advanced-stage CTE (High CTE). We analyzed WSIs using two stains: pathological AT8 immunostaining and structural Luxol fast blue, hematoxylin, & eosin (LH&E) staining. Attention-based multiple instance learning models were trained to detect subtle changes across CTE stages: NC to RHI, RHI to Low CTE, and Low CTE to High CTE. This weakly-supervised approach eliminates the need for laborious manual annotations by classifying WSIs based on slide-level diagnoses rather than pixel-level annotations. The self-learned attention scores were used to plot attention maps that highlight critical regions in the slides, facilitating the identification of relevant pathological features.

Results, Conclusions, and Discussions: Results
The model exhibited strong performance in differentiating between various stages of CTE development. For the AT8-based model, it achieved an accuracy of 0.905 and an AUC of 0.956 in distinguishing RHI from High CTE, with attention maps effectively highlighting AT8-positive regions. For the LH&E-based models, the accuracies were 0.821, 0.598, 0.615, and 0.733, and the AUCs were 0.979, 0.637, 0.663, and 0.738 for the NC to RHI, RHI to Low CTE, Low CTE to High CTE, and RHI to High CTE tasks, respectively. The attention maps successfully identified structural changes in LH&E slides that correlate with different CTE stages.
Conclusions and Discussions
Our attention-based weakly supervised learning model shows great promise for automating CTE pathology analysis in brain whole-slide images (WSIs). By reducing the need for manual annotations, this approach streamlines the diagnostic workflow and helps mitigate potential biases. The results suggest that LH&E staining, when used in conjunction with attention-based models, can effectively reveal structural changes associated with CTE. This combination offers a promising avenue for improving diagnostic accuracy and efficiency, as well as identifying new biomarkers for NDs.
Future work will involve extending our analysis to additional brain regions and stains and exploring other neurodegenerative conditions. We also plan to develop foundational models to enhance feature extraction from neuropathological images, aiming to improve the model's generalizability and interpretability. Ultimately, these efforts will contribute to better diagnostic tools and support individuals affected by NDs.

Acknowledgements (Optional): We acknowledge fundings from Boston University Kilachand Fund and the Alzheimer’s disease research center (NIA P30 AG13846).