Researchers from Harvard University and the University of Michigan Medical School are using artificial intelligence systems to assist with the diagnosis of brain tumors.
Current methods of intraoperative histopathologic diagnosis are both labor and time-intensive. They delay the decision-making process during surgery, with a delay of around 30 to 40 minutes during surgery caused using conventional imaging techniques.
Researchers have been able to speed up the process using an imaging technique called Raman Scattering (SRS). SRS can be used on fresh, unprocessed brain tissue. The label-free optical process enabling neurosurgeons to cut the time to obtain the images to around 3 minutes.
In clinical practice, neuropathologists typically diagnose brain tumors with an accuracy of between 90% and 95%, although errors in diagnosis are made and there is often variability with pathologists’ diagnoses. The researchers believe that the use of deep learning algorithms could assist with diagnoses, which could help to reduce this variability.
The use of deep learning for diagnosing brain tumors could also help diagnose patients far faster, which would improve patient outcomes. Faster detection of brain tumors means patients would need to spend less time undergoing surgical procedures and would therefore experience fewer risks from treatment.
For the study, the researchers used over 100 brain samples and developed an algorithm to detect the signs of cancerous growth. The researchers trained their ‘neural network’ using a NVIDIA GeForce GTX 1080 Graphics processing unit (GPU) with a CUDA® Deep Neural Network library (cuDNN) on the Theano deep learning framework. The system was able to categorized the samples into one of four categories.
In the tests, the deep learning algorithm was shown to have an accuracy rate of 90% on 30 tissue samples that were analyzed. However, the more tissue samples that are analyzed, the better the deep learning system will become at diagnosing tumors. The researchers are also planning on increasing the categories to eight, which will cover the majority of tumors that are encountered by neurosurgeons.
The aim is not to replace highly trained pathologists with machines, as human expertise is needed to make the final diagnosis. However, involving AI systems in the diagnostic process could help to bring more consistency and could save pathologists valuable time.
To date, the system has been tested on 370 patients, although the researchers are pushing forward to bring the total up to 500. First author of the study, Dr. Daniel Orringer, assistant professor of neurosurgery at Michigan Medicine, said “We want to bring accuracy rates up so fewer patients are misdiagnosed.” Orringer explained that now the system has been tested and has been shown to be accurate, the next stage will be a larger clinical trial; however, at this stage the system can only be used in a research setting.
The system would be particularly valuable in small or remote hospitals that do not have access to specialists. For example, while there are 1,400 hospitals in the United States that perform neurosurgery, there are only 800 hospitals in the country with board-certified neurologists.
The system would not need to be based locally, as each hospital could be connected to the system and have images subjected to analysis remotely. While the system could be used to help neurosurgeons with medical diagnoses, there is considerable potential for the system to be used to diagnose a wide range of cancers.
The study – Rapid intraoperative histology of unprocessed surgical specimens via fibre-laser-based stimulated Raman scattering microscopy – was recently published in Nature Biomedical Engineering.