Researchers wanted to see if NPH patients’ gait is influenced by their environment, and how testing methods could influence their diagnosis.
Nearly 85,000 people in the U.S. will receive a primary brain tumor diagnosis this year, which likely involves patients undergoing an invasive histopathologic assessment. Yet researchers recently developed a deep learning model that can classify what are, according to brain tumor statistics, the six most common brain tumor types. And it can do so with just one 3D MRI brain scan.
Their model is reportedly the first of its kind. Machine and deep learning techniques have been studied to improve brain tumor diagnosis in the past, but typically required manual input or use 2D sections that depict similar brain anatomy for different specific tumor types.
What’s more, these 2D approaches to brain tumor diagnosis via machine learning “require radiologists to manually delineate, or characterize, the tumor area on an MRI scan before machine processing,” which can be “tedious and labor-intensive,” according to Satrajit Chakrabarty, a doctoral student at the Mallinckrodt Institute of Radiology’s Computing Imaging Lab at Washington University School of Medicine.
Classifying Most Common Brain Tumors with a Single 3D MRI Brain Scan
In their study published in Radiology: Artificial Intelligence, Chakrabarty et al. created a large dataset of 2,105 intracranial 3D MRI brain scans from four public datasets: the Brain Tumor Image Segmentation dataset, the LGG-1p19q dataset, the Cancer Genome Atlas Glioblastoma multiforme dataset and the Cancer Genome Atlas Low Grade Glioma dataset, along with their own internal data.
These datasets were divided into a training dataset, an internal test set and an external test dataset. Their machine learning model, a convolutional neural network, was then trained in brain tumor diagnosis, that is, to distinguish between images depicting healthy tissue and images depicting tumors, as well as classify the tumors as either a high-grade glioma, low-grade glioma, brain metastasis, meningioma, pituitary adenoma or acoustic neuroma.
Using the internal test dataset, the model was able to achieve an accuracy of over 93% for the healthy tissue set and all six tumor types. Using the external dataset, which only included low- and high-grade gliomas, the model was able to achieve a 92% accuracy.
However, researchers noted that the main limitation of their study on brain tumor diagnosis and machine learning was the use of a single imaging modality, using multiple modalities would have provided additional information about the growth of gliomas and lead to a more accurate classification. But they plan to address this issue in the future, and believe their study still shows great potential for brain tumor diagnosis.
“These results suggest that deep learning is a promising approach for automated classification and evaluation of brain tumors,” according to Charkrabarty. The model, he and his team believe, is just one step closer to a completely automated workflow, eliminating the need for manual tumor segmentation altogether.
In the future, Dr. Aristeidis Soritas, a co-developer of the model, believes it can even be extended to other brain tumor types or neurological disorders.
Brain Tumor Statistics
There are believed to be over 150 documented types of brain tumors, according to brain tumor statistics from the American Association of Neurological Surgeons, but most of recent research has focused on classifying just a few of the more common ones such as gliomas or meningiomas. Chakrabarty et al.’s brain tumor diagnosis and machine learning study is one of the first, to their knowledge, to be able to classify six of the most common brain tumor types:
The National Brain Tumor Society considers glioblastomas to be the most common primary brain tumor diagnosis, accounting for 14.5% of all tumors and 48.6% of all malignant tumors. The survival rate one-year post-diagnosis is believed to be only 40%.
Low-grade gliomas are believed to account for 6.4% of all primary brain tumors and 15% of all gliomas.
An estimated 70,000—170,000 cancer patients are thought to be diagnosed with brain metastases every year, and around 100,000 of these patients will die from them each year, according to the National Brain Tumor Society. Brain metastases are thought to be 10 times more common than primary brain tumors. But diagnosing single metastases is thought to be especially challenging because they can appear nearly identical to glioblastomas on a standard MRI. In fact, up to 40% of brain metastases are thought to be misclassified.
Meningiomas are considered to be the most common type of primary central nervous system tumors, accounting for 37.6%. However, they are misdiagnosed fairly often and it can take patients several years to receive a correct diagnosis, according to the American Association for Neurological Surgeons.
Nearly 10,000 pituitary tumors are diagnosed every year in the U.S., the majority of which are believed to be pituitary adenomas, according to the American Cancer Society. However, MRI scans of deceased patients has revealed that this number may be as high as 1 out of every 4 people.
Acoustic neuromas are believed to account for 8% of all intracranial tumors. But while the tumor can often be managed, many patients continue to experience long term symptoms. For example, over 75% of patients treated for acoustic neuroma were found to have nonserviceable hearing eight years later, according to a 2015 study published in Neurosurgery.