In a major leap forward, researchers have developed Ceograph, an AI model that predicts outcomes for cancer patients by analyzing tissue samples.
The innovative discovery is described in the journal Nature Communications as a cell spatial organization-based graph convolutional network designed to analyze cell spatial organization (for example, the cell spatial distribution, morphology, proximity, and interactions) derived from pathology images. It identifies key cell spatial organization features by accurately predicting their influence on patient clinical outcomes.
The innovation is significant due to the crucial role the cell spatial organizations have on cancer biology and tumor progression. Though there have been advances in modeling the cell-cell spatial interactions and their impact on biomarker expressions, studies till now have predominantly concentrated on the spatial distance between two cells. This focus, although informative, may neglect the more important facets of cell spatial architecture, such as cell structure alignment, distinct cell types, and interactions among multiple cells.
The process of collecting the tissue samples from patients and then placing them for interpretation on slides for diagnosis is time-consuming. Such interpretations can vary among pathologists. The AI models built in the past several years can perform some aspects of a pathologist’s job, nevertheless, the existing models fall short of accurately reproducing the analysis of tissue images, the way pathologists do it.
Where was the model applied?
The scientists successfully applied this tool to three clinical scenarios using pathology slides:
- lung cancer subtype classification;
- assessment of the risk of malignant transformation in patients with oral potentially malignant disorders (OPMD); and
- predicting treatment response in lung cancer.
What makes Ceograph unique is its potential for interpretability. For instance, in the context of lung cancer subtypes, the researchers aim to see if Ceograph can capture the nuanced differences in cell organization between lung squamous cell carcinoma (SCC) and lung adenocarcinoma (ADC).
Similarly, with OPMD patients, there may be a possibility to discern if disruptions in epithelial strata structure have implications for malignant transformations.
Lastly, for lung cancer treatment response, understanding cellular features such as tumor nuclei morphology and stroma-stroma interactions could offer new discoveries into treatment sensitivities.Â
