FastGlioma Is Representative of the Role AI Can Play in Cancer Care

A novel visual foundation model may be useful in accurately and rapidly detecting glioma infiltration in fresh, unprocessed, unlabeled surgical tissue, according to study findings published in Nature.

Researchers conducted an international, prospective, multicenter cohort study to develop, train, fine-tune, and validate FastGlioma, an open-source, artificial intelligence (AI)-based bedside diagnostic tool for the detection of brain tumor infiltration in surgical tissue.

Microscopic tumor infiltration can be detected by imaging fresh tissue samples at the surgical margins of a resection cavity using stimulated Raman histology (SRH) optical imaging. SRH images are taken in 10 seconds in a lower resolution mode or in 100 seconds in a full resolution mode, allowing for rapid identification. The FastGlioma dataset was acquired from 13 medical centers and includes data from more than 3000 patients, spanning the diagnostic spectrum of central nervous system tumors and human cancers.

Fine-tuning strategies allow generalization to wider populations. The tumor infiltration dataset used for fine-tuning was annotated by 3 neuropathologists and each SRH image was ranked on a scale ranging from 0 to 3:

• Score of 0: normal brain tissue with no tumor;
• Score of 1: atypical cells with possible tumor;
• Score of 2: sparse tumor infiltration; and,
• Score of 3: dense tumor infiltration.

Scores were confirmed with hematoxylin and eosin tissue staining and tumor-specific immunohistochemistry (ie, isocitrate dehydrogenase [IDH]-1 or -2, p53). These images were then used for comparison when rapidly identifying dense tumor infiltrations.

The model was tested in a population sample recruited from the medical centers at University of California San Francisco, New York University, and Medical University of Vienna. Adults with diffuse gliomas who were undergoing tumor resection were eligible for inclusion.

A total of 220 patients were included in the study, resulting in 767 IDH-mutant and 659 IDH-wild type specimens. The average area under the receiver operator characteristic curve (AUROC) for differentiating the degrees of diffuse glioma infiltration was 92.1 ± 0.9%. Normalized infiltration scores demonstrated a strong correlation with ground-truth ordinal labels (correlation coefficient, P =.77; 95% CI, 0.74-0.78; P =.00). A less than 1% reduction in prediction accuracy was exhibited by the use of low vs full resolution imaging.

To evaluate the feasibility and safety of FastGlioma as a surgical adjunct, an interventional clinical trial in which surgical resections are guided by FastGlioma predictions was simulated in a subset of 129 patients, resulting in 624 specimens.

FastGlioma outperformed image- and fluorescence-guided methods for tumor infiltration detection. FastGlioma achieved an AUROC of 98.1% vs 76.3% for FLAIR positivity, 71.8% for contrast engagement, and 89.0% for 5-aminolevulinic acid fluorescence.

A smaller proportion of patients in the FastGlioma vs surgical adjuncts arm had at least 1 high-risk tumor missed (3.8% vs 24.0%).

“FastGlioma has the potential for immediate clinical impact on improving the comprehensive management of patients with diffuse glioma,” the study authors concluded.

Disclosure: Multiple study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.