An artificial intelligence (AI)-based tool can predict local failure (LF) after radiation therapy for brain metastases (BM) using the baseline treatment-planning MRI, a new study suggests.
The tool was 83% accurate in predicting treatment failure, while oncologists accurately predict treatment failure about 65% of the time using conventional methods, according to York University, Toronto, Ontario, Canada.
The study shows that AI-detected perilesional characteristics that are often overlooked on treatment-planning MRI scans can be used to predict LF after radiation.
Ali Sadeghi-Naini, PhD
Up to 30% of BM do not respond to stereotactic radiation therapy, “and it takes months before oncologists can assess the outcome for a tumor using standard follow-up (serial) imaging,” study author Ali Sadeghi-Naini, PhD, professor of electrical engineering and computer science at York University, Toronto, Ontario, Canada, told Medscape Medical News.
However, he said, “MRIs potentially carry much more information than is currently exploited in standard practice. AI models can help to capture and utilize such information and possibly serve as valuable clinical decision-support tools in BM management.”
The study was published online November 4 in the IEEE Journal of Translational Engineering in Health and Medicine.
Focusing Attention
To facilitate early prediction of LF, the researchers began developing AI and machine learning tools about 5 years ago, said Sadeghi-Naini. Their first study showed the feasibility of predicting LF by comparing certain features of the first follow-up scan with the pretreatment scan. Their second study showed that the pretreatment scan alone could be used for prediction. Their next study showed that a deep learning approach using 2D MRI slices and clinical attributes for prediction could outperform the assessment of clinical variables alone.
The current study analyzed the entire 3D MRI volume using deep learning, with “attention mechanisms” trained to focus on certain aspects of the MRI. The investigators developed the latest model based on data from 124 patients with 156 lesions. The data were randomly divided into a training set (99 patients with 116 lesions) and a test set (25 patients with 40 lesions) that was used for independent evaluation. Ten patients with 15 lesions were randomly selected from the training set as the validation set for optimizing the model.
Patients had an average age of 62 years, and 60% were women. The average tumor size was 2 cm, and the average Graded Prognostic Assessment was 2.2.
The researchers taught the AI model to focus its attention on “sophisticated patterns in the MRI intensity that can describe the heterogeneity in different areas surrounding the tumor,” explained Sadeghi-Naini. “Capturing and quantifying these patterns or features is not often possible with the human eye alone.”
This latest model outperformed previous versions and demonstrated the best performance in terms of area under the curve and F1-score (a machine learning evaluation metric that measures a model’s accuracy on a data set).
“The visualization results show the importance of perilesional characteristics on treatment-planning MRI in predicting local outcome after radiotherapy and the feasibility of early prediction of radiotherapy outcome for BM using only the features extracted from multimodal MRI volumes,” wrote the authors.
Next, the team will evaluate the model in larger, multi-institution data sets and, if warranted, design clinical trials to further test the model’s performance. “We expect to reach that point in 3-5 years,” Sadeghi-Naini concluded.
Clinically Relevant?
Commenting on the study for Medscape, Mike Y. Chen, MD, PhD, associate professor and director of the neurosurgical oncology fellowship program at City of Hope (CoH) in Duarte, California, said, “Overall, it is a promising approach that will be essential in the future. The amount of data a radiologist or clinician has to review is enormous. For example, at CoH, the basic brain MRI for tumor evaluation consists of eight sequences totaling 617 images. More specialized exams easily surpass 1000 images.”
Dr Mike Chen
Most clinicians have about a half hour to compare a new MRI with at least one previous image, which is not enough time to examine the nuances carefully, said Chen. “A computer-aided analysis can be a tremendous asset, [and] AI can discover new details on the scans that humans have overlooked that may be clinically relevant.”
Chen raised caveats about the study and the approach, however. The quantity and quality of the training material is “critically important” to avoid “junk in, junk out,” he said. A dataset of 100 patients “is respectable, but not nearly large enough.”
The quality of the data is also “limited,” he added. Details not provided include the definition of LF (eg, a 20% increase in size over 8 weeks post-treatment) and whether the investigators accounted for radiation necrosis, which can make lesions appear larger after treatment, when in fact the tumor is “dead.”
“The larger issue is really that of clinical relevance,” he said. “How will results provided by the AI change a physician’s decisions and improve clinical care? I suspect that providing this burden of proof will be as challenging as developing the technology itself.”
The study was supported by the Natural Sciences and Engineering Research Council of Canada, the Lotte and John Hecht Memorial Foundation, and the Terry Fox Foundation. Sadeghi-Naini and Chen reported no conflicts of interest.
IEEE J Transl Eng Health Med. Published online November 4, 2022. Full text.
Follow Marilynn Larkin on Twitter: @MarilynnL.
For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube.
Source: Read Full Article
