Fig. 3: Interpretable AI-guided diagnosis of spine tumors across international institutions.
From: AI-driven label-free Raman spectromics for intraoperative spinal tumor assessment
Representative specimens from four common spine tumor types are shown with their corresponding confidence scores (maximum 100): meningioma (100), schwannoma (99), ependymoma (81), and metastasis (97). The model generates color-coded and interpretable visualization overlays that highlight diagnostically relevant regions specific to each tumor type. Below each prediction, characteristic histological features are displayed in magnified SRH patches, revealing distinctive patterns such as psammoma bodies in meningioma, fascicle formation in schwannoma, ependymal rosettes in ependymoma, and nuclear pleomorphism in metastasis. SpineXtract maintains consistent performance across institutions (New York, Michigan, Vienna), diverse anatomical locations (intradural extramedullary, intramedullary and extradural) and histological presentations, providing surgeons with immediate visual guidance during intraoperative decision-making.
