Supplementary Table S1. Description of the SRH-based diagnostic accuracy of the first model in case of neuropathological diagnosed non-tumorous tissue from open resection procedures since the stereotactic-guided biopsy dataset contained only a few cases due to IRB protocol restrictions and the common non-invasive diagnostic differentiation by means of 18F-FET-PET CT/MRI to distinguish between tumorous and non-tumorous tissue, e.g., in recurrent tumor cases. Non-tumorous tissue samples of the brain summarized here were obtained non- or purposely from surrounding white or gray matter adjacent to the tumor during open tumor en bloc resections. After resection, these tissue parts were extracted from the whole tissue sample for further analysis. Further cases comprised space-occupying suspected tumorous lesions or post-radiotherapy lesions (e.g., pseudoprogression or radiogenic necrosis) due to immense edematous effect. The final FFPE-based diagnosis revealed non-tumorous tissue.
ARTICLE ABSTRACT
Recent artificial intelligence algorithms aided intraoperative decision-making via stimulated Raman histology (SRH) during craniotomy. This study assesses deep learning algorithms for rapid intraoperative diagnosis from SRH images in small stereotactic-guided brain biopsies. It defines a minimum tissue sample size threshold to ensure diagnostic accuracy.
A prospective single-center study examined 121 SRH images from 84 patients with unclear intracranial lesions undergoing stereotactic brain biopsy. Unprocessed, label-free samples were imaged using a portable fiber laser Raman scattering microscope. Three deep learning models were tested to (i) identify tumorous/nontumorous tissue as qualitative biopsy control; (ii) subclassify into high-grade glioma (central nervous system World Health Organization grade 4), diffuse low-grade glioma (central nervous system World Health Organization grades 2–3), metastases, lymphoma, or gliosis; and (iii) molecularly subtype IDH and 1p/19q statuses of adult-type diffuse gliomas. Model predictions were evaluated against frozen section analysis and final neuropathologic diagnoses.
The first model identified tumorous/nontumorous tissue with 91.7% accuracy. Sample size on slides impacted accuracy in brain tumor subclassification (81.6%, κ = 0.72 frozen section; 73.9%, κ = 0.61 second model), with SRH images being smaller than hematoxylin and eosin images (4.1 ± 2.5 mm2 vs. 16.7 ± 8.2 mm2, P < 0.001). SRH images with more than 140 high-quality patches and a mean squeezed sample of 5.26 mm2 yielded 89.5% accuracy in subclassification and 93.9% in molecular subtyping of adult-type diffuse gliomas.
Artificial intelligence–based SRH image analysis is non-inferior to frozen section analysis in detecting and subclassifying brain tumors during small stereotactic-guided biopsies once a critical squeezed sample size is reached. Beyond frozen section analysis, it enables valid molecular glioma subtyping, allowing faster treatment decisions in the future; however, refinement is needed for long-term application.
