News | Analytics Software | October 25, 2016

Computer Program Beats Physicians at Brain Cancer Diagnoses

New research could speed identification of recurrent tumors, eliminate costly and risky brain biopsies

Case Western Reserve University study, machine learning, MRI, brain cancer diagnoses, radiomics

MRI scans of patients with radiation necrosis (above) and cancer recurrence (below) are shown in the left column. Close-ups in the center column show the regions are indistinguishable on routine scans. Radiomic descriptors unearth subtle differences showing radiation necrosis, in the upper right panel, has less heterogeneity, shown in blue, compared to cancer recurrence, in the lower right, which has a much higher degree of heterogeneity, shown in red. Credit: Pallavi Tiwari

October 25, 2016 — Computer programs have defeated humans in Jeopardy!, chess and Go. Now a program developed at Case Western Reserve University has outperformed physicians on a more serious matter.

The program was nearly twice as accurate as two neuroradiologists in determining whether abnormal tissue seen on magnetic resonance images (MRI) were dead brain cells caused by radiation, called radiation necrosis, or if brain cancer had returned.

The direct comparison is part of a feasibility study published in the American Journal of Neuroradiology.

“One of the biggest challenges with the evaluation of brain tumor treatment is distinguishing between the confounding effects of radiation and cancer recurrence,” said Pallavi Tiwari, assistant professor of biomedical engineering at Case Western Reserve and leader of the study. “On an MRI, they look very similar.”

But treatments for radiation necrosis and cancer recurrence are far different. Quick identification can help speed prognosis, therapy and improve patient outcomes, the researchers say.

With further confirmation of its accuracy, radiologists using their expertise and the program may eliminate unnecessary and costly biopsies, Tiwari said. Brain biopsies are currently the only definitive test but are highly invasive and risky, causing considerable morbidity and mortality.

To develop the program, the researchers employed machine learning algorithms in conjunction with radiomics, the term used for features extracted from images using computer algorithms. The engineers, scientists and physicians trained the computer to identify radiomic features that discriminate between brain cancer and radiation necrosis, using routine follow-up MRI scans from 43 patients. The images were all from University Hospitals Case Medical Center.

The team then developed algorithms to find the most discriminating radiomic features — in this case, textures that can’t be seen by simply eyeballing the images.

“What the algorithms see that the radiologists don’t are the subtle differences in quantitative measurements of tumor heterogeneity and breakdown in microarchitecture on MRI, which are higher for tumor recurrence,” said Tiwari, who was appointed to the Department of Biomedical Engineering by the Case Western Reserve School of Medicine.

More specifically, while the physicians use the intensity of pixels on MRI scans as a guide, the computer looks at the edges of each pixel, explained Anant Madabhushi, F. Alex Nason professor II of biomedical engineering at Case Western Reserve, and study co-author.

“If the edges all point to the same direction, the architecture is preserved,” said Madabhushi, who also directs the Center of Computational Imaging and Personalized Diagnostics at CWRU. “If they point in different directions, the architecture is disrupted — the entropy, or disorder, and heterogeneity are higher. “

In the direct comparison, two physicians and the computer program analyzed MRI scans from 15 patients from University of Texas Southwest Medical Center. One neuroradiologist diagnosed seven patients correctly, and the second physician correctly diagnosed eight patients. The computer program was correct on 12 of the 15.

Tiwari and Madabhushi do not expect the computer program would be used alone, but as a decision support to assist neuroradiologists in improving their confidence in identifying a suspicious lesion as radiation necrosis or cancer recurrence.

Next, the researchers are seeking to validate the algorithms’ accuracy using a much larger collection of images from across different sites.

For more information: www.ajnr.org

Related Content

RADLogics AI-Powered solution in use: chest X-ray of COVID-19 positive case with heatmap key image.

RADLogics AI-Powered solution in use: chest X-ray of COVID-19 positive case with heatmap key image.

News | Artificial Intelligence | September 23, 2020
September 23, 2020 — RADLogics
The cartilage in this MRI scan of a knee is colorized to show greater contrast between shades of gray.

The cartilage in this MRI scan of a knee is colorized to show greater contrast between shades of gray. Image courtesy of Kundu et al. (2020) PNAS

News | Artificial Intelligence | September 22, 2020
September 22, 2020 — Researchers at the University of Pitts...
New research from King's College London has found that COVID-19 may be diagnosed on the same emergency scans intended to diagnose stroke.

Canon Medical Systems

News | Cardiac Imaging | September 22, 2020
September 22, 2020 — New research from King's College London has
Philips Azurion Lung Edition supports high precision diagnosis and minimally invasive therapy in one room
News | Lung Imaging | September 21, 2020
September 21, 2020 — Philips introduced...
According to a new report published by P&S Intelligence, the global radiotherapy market is expected to expand from $7.2M in 2019 to $17M by 2030.

Image courtesy of Accuray

Feature | Radiation Therapy | September 21, 2020 | By Melinda Taschetta-Millane
According to a...
According to Philips, MR-STAT is a major shift in MRI, relying on a new, smart acquisition scheme and machine-assisted reconstruction. It delivers multiple quantitative MR parameters in a single fast scan, and represents a significant advance in MR tissue classification, fueling big data algorithms and AI-enabled integrated diagnostic solutions.

Image courtesy of Philips Healthcare

Feature | Magnetic Resonance Imaging (MRI) | September 21, 2020 | By Melinda Taschetta-Millane
A new report,...
Of all the buzzwords one would have guessed would dominate 2020, few expected it to be “virtual”. We have been virtualizing various aspects of our lives for many years, but the circumstances of this one has moved almost all of our lives into the virtual realm.

Getty Images

Feature | Radiology Education | September 18, 2020 | By Jef Williams
Of all the buzzwords one would have guessed would dominate 2020, few expected it to be “virtual”.
Indeterminate lesion on PET/CT classified by PET/MRI for 53-y-old man with lung cancer.

Indeterminate lesion on PET/CT classified by PET/MRI for 53-y-old man with lung cancer. Contrast-enhanced CT (A), PET (B), and fused 18F-FDG PET/CT (C) images are displayed in comparison with contrast-enhanced T1-weighted MRI (D), PET, and fused 18F-FDG PET/MRI (F) images. In CT (A), hyperdense, subcentimeter liver lesion (arrows) in segment VII is suggestive of transient hepatic attenuation difference or small hemangioma. As malignancy cannot be excluded, it needs further investigation. On PET/MRI, lesion is clearly classified as metastasis because of contrast enhancement and tracer uptake due to later acquisition time point. Follow-up CT confirmed diagnosis after 78 d. Images created by Ole Martin, University Dusseldorf, Medical Faculty and Benedikt Schaarschmidt, University Hospital Essen.

News | PET-MRI | September 18, 2020
September 18, 2020 — A single-center observational study of more than 1,000 oncological examinations has demonstrated