There are "twin" brain tumors that look alike even though their causes, progression, and treatments are entirely different. Because cancer cell shapes and tissue structures are so similar, misdiagnoses have been frequent. U.S. scientists have developed artificial intelligence (AI) that accurately distinguishes the two brain tumors.
Harvard Medical School researchers said they developed an AI tool, "PICTURE," that distinguishes glioblastoma from primary central nervous system lymphoma (PCNSL) from camera images with 98% accuracy, according to a paper published in the international journal Nature Communications on the 29th (local time).
Glioblastoma is the most common and aggressive tumor arising from brain cells. By contrast, PCNSL arises from immune cells but is often misdiagnosed as glioblastoma. Treatment strategies are the opposite. In glioblastoma, the tumor should be resected as much as possible, whereas in PCNSL, radiation and chemotherapy take precedence. Accurately distinguishing the two cancers therefore determines the success or failure of treatment.
The researchers noted that PICTURE's strength is that it can be used immediately during surgery. Typically, tissue biopsy for cancer diagnosis involves excising tumor tissue, rapidly freezing it in liquid nitrogen, and examining it under a microscope. This approach can distort cell morphology and reduce accuracy. In fact, in about 1 out of 20 cases, the result changes in the definitive diagnosis a few days later. It also takes a long time.
The team overcame these limitations with AI. Kun-Hsing Yu, a Harvard Medical School professor who led the study, said, "Our model clearly distinguishes tumors with similar features to reduce diagnostic errors and helps select the optimal treatment tailored to the patient."
The AI proved its accuracy in real patient diagnoses. In validation at five hospitals in the United States, Europe, and Asia, PICTURE distinguished glioblastoma and PCNSL with over 98% accuracy. It also identified 67 other central nervous system tumors, and it correctly classified cases that pathologists misdiagnosed by as much as 38%, outperforming clinicians and existing AI.
Unlike existing AI, including the generative AI ChatGPT, PICTURE also does not pretend to know what it does not. When it encounters an unfamiliar tumor, it does not force an answer; instead, it displays a "review needed" signal so that an expert can reexamine it.
However, a limitation is that most of the biospecimens used for AI training came from white patients. The researchers said they will verify accuracy across diverse racial groups and expand the scope of research to other cancers. They also plan to integrate genetic and molecular analysis data to attempt deeper analyses.
References
Nature Communications (2025), DOI: www.doi.org/10.1038/s41467-025-64249-6