MRI images enable ‘virtual’ biopsies to speed analysis of brain tumors
A novel deep learning technique can accurately identify genetic mutations in tumors that originate in the brain’s supportive tissues MRI images.
Researchers say that this technology can speed up diagnoses and enable patients to receive more personalized treatment regimens for the condition, known as gliomas. These are a heterogeneous group of primary tumors with variable imaging characteristics, responses to therapy, clinical courses and prognoses. This diversity is attributed in part to the many genetic and epigenetic mutations in the tumors.
The knowledge of a tumor’s genetic information is needed to accurately monitor patients and guide a patient’s therapy. However, currently information regarding the mutations of gliomas is based on an analysis of tumor tissue obtained during an operation. These are often limited to the easily accessible area of the tumor, and that may potentially miss other, different variations within the tumor. This testing can also be expensive, and the results can take weeks, delaying treatment.
Noninvasive imaging can provide complementary insight into the tumor and act as a “virtual” biopsy. It can also assess the entire tumor, not just the area that was biopsied, to better classify the tumors.
Prior machine learning approaches linking images to the genetic alterations in gliomas relied on human derived feature extraction, such as textural analysis.
The researchers, from the University of California Irvine, Columbia University and elsewhere developed a deep learning algorithm implemented with convolutional neural networks to classify three genetic variations of gliomas. They used MR imaging data and molecular information retrospectively obtained from the Cancer Imaging Archives for 259 patients with low or high grade gliomas.
The algorithm correctly identified one mutation (isocitrate dehydrogenase 1 (IDH1) mutation status) with 94 percent accuracy, another (1p/19q codeletion) with 92 percent accuracy and the third (O6-methylguanine-DNA methyltransferase (MGMT) promotor methylation status) with 83 percent accuracy.
The algorithm also identified distinctive imaging features for each genetic mutation, such as the definition of tumor margins, the extent of edema, necrosis and textual features.
Notably, the technology used end-to-end machine learning, automatically identifying patterns in the imaging sets, combining both tumor detection, tissue segmentation and mutation classification into one algorithm.
“The current study is the first to demonstrate the feasibility of a single neural network architecture to simultaneously predict the status of multiple different mutations (IDH1 status, 1p/19q codeletion, MGMT promoter methylation) with minimal preprocessing in an efficient, fully automated approach,” the study authors say.
The study was recently published in the American Journal of Neuroradiology.
“The results of our study show the feasibility of a deep-learning [convolutional neural network] approach for the accurate classification of individual genetic mutations of both low- and high-grade gliomas. Furthermore, we demonstrate that…neural networks are capable of learning key imaging components without prior feature selection or human directed training,” the study authors conclude.
The AI developed for this study may be one of the first applications of deep learning to be used in clinical care by the University of California, Irvine and UCI Health System’s new Center for Artificial Intelligence in Diagnostic Medicine, a multispecialty program to integrate AI technology into routine clinical practice throughout the Health System.