A new AI framework reads MRI scans and tumor biopsy slides together, aligning two very different views of the same brain tumor to sharpen glioma diagnosis.
The system, called CrossScale-GLIO, converts an MRI into a graph of tumor habitat regions and converts a microscope slide into a graph of cell neighborhoods, then matches the two graphs using an optimal transport algorithm guided by diagnostic text. Tested on paired and external glioma datasets, it predicted tumor subtype with a macro-F1 score of 0.789 and identified three genetic markers used in treatment planning, IDH mutation, 1p/19q co-deletion, and MGMT methylation status, with AUROC scores of 0.934, 0.884, and 0.802. It also matched MRI scans to their correct tissue-sample counterparts in a retrieval test, though only about 28 percent of the time on the first guess and 62 percent within the top five guesses.
The more convincing evidence is in the ablation tests, not the headline accuracy numbers. Deleting the single highest weighted MRI to tissue link cut classification confidence by 0.184, four times the drop from a random deletion, and scrambling the graph's structure while preserving its size reduced both subtype accuracy and retrieval performance. That suggests the model is learning real correspondence between radiology and pathology features, which is usually the weak point of AI trained on small paired imaging and genomic datasets.
Still, this is one retrospective study on modest cohorts with no reported patient counts, and the finding that pathologists rated 81.2 percent of pairings as biologically plausible is a soft, subjective benchmark, a long way from anything approaching clinical validation.