A Harvard Medical School-led research team has developed an AI tool called PICTURE that accurately differentiates between glioblastoma and primary central nervous system lymphoma (PCNSL), two brain cancers that often appear similar under the microscope.

The AI tool is intended to support neuro-oncology diagnostics, reduce errors, and serve as an educational resource for training future pathologists.

PICTURE's consistent performance across various tissue types, including frozen and fixed samples, underscores its robustness and potential for broad clinical application.

Future developments include expanding the AI's diagnostic scope, integrating genetic insights, and validating its effectiveness in diverse patient populations.

Accurate intraoperative diagnosis is critical for guiding treatment, and PICTURE's real-time analysis helps minimize errors, ensuring more appropriate and timely interventions.

In challenging cases where pathologists may misdiagnose up to 38% of the time, PICTURE provides a valuable decision support tool that enhances diagnostic accuracy.

PICTURE achieved over 98% accuracy in distinguishing these tumors, significantly reducing diagnostic errors and supporting real-time decision-making during brain surgery.

The AI system utilizes advanced pathology image analysis with an 'uncertainty detector' to identify cases where its confidence is low, prompting human review and enhancing diagnostic reliability.

Designed for use during surgery, PICTURE provides immediate insights that can guide treatment choices, such as tumor removal versus radiation or chemotherapy, thereby improving surgical outcomes.

Supported by the NIH and tested across five international hospitals with over 2,100 pathology slides, PICTURE outperforms both human pathologists and other AI models, especially in challenging cases.

The model is publicly available for further research and development, with the aim of democratizing access to specialized neuropathology and supporting less experienced clinicians.

While most samples tested were from white patients, further validation across diverse populations is necessary, and future updates may expand its capabilities to other tumor types and incorporate genetic data.