Your Daily digest for NIH Director's Blog Daily Digest (Unofficial)

[Article Digests for Psychology & Social Work]

NIH Director's Blog Daily Digest (Unofficial)

 

(https://directorsblog.nih.gov/2024/12/12/ai-system-has-potential-to-differentiate-brain-cancer-from-healthy-tissue-during-surgery-within-seconds/) AI System Has Potential to Differentiate Brain Cancer from Healthy Tissue During Surgery within Seconds
Dec 12th 2024, 09:00

In a study, researchers trained an AI tool called FastGlioma to quickly distinguish cancerous from healthy tissue during surgery to remove brain tumors. Credit: Donny Bliss/NIH, fastglioma.mlins.org

In people with brain tumors known as diffuse gliomas, cancerous cells often spread and invade nearby tissue to mix with healthy cells. As a result, it can be challenging for neurosurgeons to differentiate cancerous from healthy tissue during surgery as is required to safely remove as much of the cancer as possible. Many patients with glioma are found to have residual tumor after surgery, which can mean additional surgeries, earlier recurrence, and decreased survival. But research is showing that artificial intelligence (AI) tools could enable doctors to not only (https://directorsblog.nih.gov/2024/10/03/ai-model-takes-new-approach-to-performing-diagnostic-tasks-in-multiple-cancer-types/) predict if a cancer will respond to treatment, but also to differentiate cancerous from healthy tissue rapidly enough to guide more brain surgeries in real time.

In one promising example of this, an NIH-supported study in (https://www.nature.com/articles/s41586-024-08169-3) Nature recently reported the development of an open-source, AI-based diagnostic system that can determine in just 10 seconds if part of a cancerous brain tumor that could be removed still remains. The new system, called FastGlioma, combines rapid, user-friendly, optical microscopy with AI models trained on diverse data, including over 11,000 surgical specimens and 4 million microscopy images, to give surgeons needed answers very quickly.

Today, neurosurgical teams locate residual tumor during surgery guided by MRI or fluorescent imaging. The research team for this study—led by (https://www.uofmhealth.org/profile/34302/todd-charles-hollon-md) Todd Hollon, University of Michigan Health, Ann Arbor, and (https://www.ucsfhealth.org/providers/dr-shawn-hervey-jumper) Shawn Hervey-Jumper, University of California, San Francisco—reports that the new system significantly outperforms current methods for identifying tumor remains, working faster and more accurately.

FastGlioma is built on what’s known as a visual foundation model. These models are developed by training computers on massive amounts of data that can be adapted to a wide range of tasks. Such AI models have great promise for solving many clinical problems in medicine. They can not only classify images, but also act as chatbots, reply to emails, and generate images from text descriptions. According to the researchers, FastGlioma combines visual foundation model training with fine-tuning strategies that allow it to work across patients with different demographics receiving care in different health care systems for different brain tumor diagnoses and with minimal supervised training by humans.

The researchers first fine-tuned FastGlioma in an international study. They examined more than 1,000 brain tissue samples from 220 patients with diffuse gliomas who underwent brain surgery. Overall, FastGlioma detected and quantified how much tumor remained with an average accuracy of approximately 92%.

Next, to test how well the system could differentiate healthy and cancerous tissue in the operating room compared to current standard methods, the researchers conducted a simulated trial in 129 people with diffuse glioma who had surgery. While FastGlioma isn’t approved for use and wasn’t used to make treatment decisions, the trial allowed the study team to compare FastGlioma’s predictions to those that doctors got using standard methods to determine the tool’s safety and efficacy. Altogether, they had 624 tumor specimens with matched predictions about whether there was cancer or not based on both FastGlioma and standard imaging methods for comparison.

The simulated trial found that FastGlioma worked better to distinguish cancer from non-cancer, with an accuracy of 98% compared to about 72% to 89% for other standard methods. Finally, the researchers looked to see how many patients had high-risk tumors missed with FastGlioma compared to standard imaging methods. Overall, they reported that the AI technology missed high-risk, residual tumor tissue in 3.8% of the patients (5 of 129), compared to a 24% miss rate (31 of 129) for the standard methods. The researchers hope to conduct a future trial in which FastGlioma will be tested further by allowing surgeons to use it to guide decision-making in the operating room.  

The presence of residual tumor tissue following surgery is a significant and costly public health problem in the U.S. and around the world, for brain cancers and other solid cancers alike. The research team reports that FastGlioma can already accurately detect residual tumor in many other brain cancer types, including both adult and childhood brain cancers, suggesting it has potential to one day serve as a general-purpose tool for guiding brain tumor surgeries. The researchers also plan to explore the system’s application to other cancers, including lung, prostate, breast, and head and neck cancers. Through this kind of work, the researchers hope this tool and others like it can help unlock the potential of AI for improving cancer care in the years ahead.

Reference:

Kondepudi A, et al. (https://pubmed.ncbi.nlm.nih.gov/39537921/) Foundation models for fast, label-free detection of glioma infiltration. Nature. DOI: 10.1038/s41586-024-08169-3 (2024).

An interactive demo of FastGlioma is available (https://fastglioma.mlins.org/) here.

NIH Support: National Institute of Neurological Disorders and Stroke, National Cancer Institute, National Institute of General Medical Sciences

Forwarded by:
Michael Reeder LCPC
Baltimore, MD

This information is taken from free public RSS feeds published by each organization for the purpose of public distribution. Readers are linked back to the article content on each organization's website. This email is an unaffiliated unofficial redistribution of this freely provided content from the publishers. 

 

(#) unsubscribe from this feed
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.clinicians-exchange.org/pipermail/article-digests-clinicians-exchange.org/attachments/20241213/f66e466f/attachment.htm>