Artificial Intelligence
AI Deep Learning Predicts Brain Tumor Growth
AI and MRI images advance personalized medicine for deadly brain cancer.
Posted January 23, 2023 Reviewed by Davia Sills
Key points
- Scientists used actual imaging data from five patients with untreated glioblastoma tumors.
- This offered them a rare opportunity to study the natural progression of the disease.
- They were able to create an AI deep learning model with the data to predict how glioblastoma advances if untreated.
Artificial intelligence (AI) deep learning combined with patient imaging data is opening a new world of possibilities for medical clinicians and researchers, especially in neuroscience. A new study published in the Journal of Theoretical Biology demonstrates how AI deep learning can predict brain tumor progression for glioblastoma from medical images to accelerate precision medicine.
“Our work provides a new, easily generalizable method for the estimation of patient-specific tumor parameters, which can be built upon to aid physicians in designing personalized treatments,” wrote study researchers affiliated with the University of Waterloo, the University of Toronto, and St. Michael’s Hospital in Toronto.
The most common cancer that originates in the brain is glioblastoma, also known as glioblastoma multiforme (GBM). Glioblastoma is a brain tumor that arises from the glial cells that surround and support neurons. It is one of the most treatment-resistant, complex, and deadly forms of cancers. Glioblastoma is incurable and the most common primary brain tumor in adults.
The most common areas where glioblastoma develops are in the frontal and temporal lobe, and rarely in the spinal cord or brain stem according to the University of Texas MD Anderson Cancer Center. The World Health Organization Classification of Tumors of the Central Nervous System categorizes glioblastoma in “Grade 4.” It is an incurable disease with a poor overall survival and a high rate of recurrence, according to the National Center for Biotechnology Information. Cancer of the brain and nervous system accounted for over 250,000 deaths worldwide in 2020, according to Global Cancer Statistics (GLOBOCAN).
What sets this study apart from other AI oncological studies is that the scientists had the rare opportunity to use actual imaging data for an uncommon cancer for five untreated tumors. The study participants were glioblastoma patients who opted not to receive any intervention or treatment at the time of the study.
“As our study aims to characterize the natural progression of the tumor, a requirement for our data was that no anticancer treatment was performed between the imaging times for each patient,” the researchers wrote.
The researchers set out to create a deep learning model that can accurately estimate patient-specific tumor progression using the Proliferation-Invasion (PI) model, a mathematical model often used to characterize glioblastoma progression.
Developing high-performance AI algorithms require massive datasets to train model so they can learn the features from the data. Finding large training datasets for uncommon cancers and rare diseases is challenging, especially where there are poor prognoses and short life expectancies. For this study, scientists used two sets of actual tumor progression data of magnetic resonance imaging (MRI) images taken several months apart from five patients diagnosed with glioblastoma.
The researchers used an open source toolkit that applies AI machine learning for spotting tumor boundaries called Federated Tumor Segmentation (FeTS) initiative software that was developed and maintained by the Center for Biomedical Image Computing and Analytics at the University of Pennsylvania. This AI toolkit uses deep learning algorithms for its segmentation of tumors into four tissue categories consisting of non-tumoral, enhancing proliferative, peritumoral edema, and necrotic. The specificity of the classification offers advantages over a binary classification of tumor/non-tumor that give the researchers richer insights for calculating the tumor cell density.
The AI deep learning model needed tumor cell density at each imaging time for input data. A popular method of obtaining tumor cell density from imaging depends on the ADC image. The apparent diffusion coefficient (ADC) measures the magnitude of diffusion of water molecules within tissue that is typically calculated from MRI with diffusion-weighted imaging. Prior unrelated studies show that a low ADC value corresponds to restricted fluid motion and high tumor cellularity—the number of cells in a tumor.
The AI deep segmentation enabled the scientists to derive tumor cellularity from multi-sequence MRI with a more sophisticated incorporation of the ADC data. To test their deep learning model, the scientists used synthetic tumors created computationally.
“Since the true parameter values and growth curve are known, we can compare the predictions of the deep learning,” the researchers wrote.
The results were that the AI predicted cellularity with a high degree of accuracy, according to the scientists. The researchers applied the AI deep learning model to the patient dataset to create a model that can predict how glioblastoma progresses if untreated, a valuable research insight. With this proof-of-concept, the next steps are to extend the algorithm to predict glioblastoma progression that includes cancer treatment, where there are much more big data sets available to train the model.
“This method allows for more accurate and personalized predictions of tumor growth and potential treatment response than were otherwise possible, which clearly has theoretical and clinical utility,” the researchers reported. “A particular advantage of this method is that it requires data only from the patient about whom predictions are being made. This sidesteps the major hurdle of requiring a large existing dataset which commonly plagues the application of machine learning models to problems in medicine.”
Copyright © 2023 Cami Rosso. All rights reserved.
