Abstract 5491: AI foundation model for single cell annotation from conventional histopathology images of cancer.

Key Points

• The study aims to develop an AI foundation model for accurate single-cell identification from conventional histopathology images.
• Developed a multimodal AI model using spatial proteomic and transcriptomic data.
• Trained the model on multiplexed immunohistochemistry and transcriptomic slides from colorectal cancer.
• Used a Multi-Layer Perceptron neural network for supervised classification of single cells.
• Achieved an overall accuracy of 87.1% on training data with a macro-average AUROC of 96.1%.
• Validation on independent CRC samples yielded consistent accuracy of 87.2% and 95% AUROC.
• Provided insights into immune-tumor interactions and cellular distributions within the tumor microenvironment.