An artificial intelligence (AI) program has been created to identify minor brain irregularities by researchers at University College London. To create the algorithm, the research collected MRI images of more than 22 patients from 1000 international epilepsy institutions.
The enhancement is considered crucial, especially for finding anomalies.
An artificial intelligence (AI) program has been created by a group of researchers from University College London to identify drug-resistant focal cortical dysplasia (FCD), a minor brain defect that causes epileptic seizures. This is an encouraging development for researchers working to detect and treat epilepsy in its early stages.
The Multicenter Epilepsy Lesion Detection project (MELD) collected MRI scans of more than 22 patients from 1000 international epilepsy centers to create the algorithm that identifies anomalies in cases of drug-resistant focal cortical dysplasia (FCD), the primary cause of epilepsy.
Brain regions known as FCDs have evolved inappropriately, resulting in drug-resistant epilepsy. Surgery is usually required for treatment. However, these lesions can be difficult to detect using this technology, as FCDs appear normal on MRI images.
To create the system, the work team used approximately 300.000 locations throughout the brain to measure cortical features from MRI scans. The algorithm was then trained on cases that professional radiologists classified as having FCD or having a healthy brain.
The findings showed that on average 67 percent of cases in the cohort had their FCD correctly identified by the algorithm (538 participants).
Radiologists failed to discover the anomaly in 178 of the subjects, according to previous research that failed to reveal it. On the other hand, in 63% of these samples, the MELD algorithm was successful in identifying FCD.
Advances in Epilepsy Treatment
This progression is critical, especially for identifying the abnormality. “We are focused on developing an interpretable AI algorithm that can assist doctors in making decisions. According to Mathilde Ripart, one of the paper's first authors, an important step in this process was to show clinicians how the MELD algorithm generates its predictions. ”
Senior author of the study, Dr. “This algorithm could help find more of these hidden lesions in children and adults with epilepsy, and enable more epilepsy patients to be considered for neurosurgery that can be treated,” says Konrad Wagstyl. epilepsy and improve their cognitive development.” In the UK, he continued, “around 440 children a year could gain from epilepsy surgery.”
Summary of Epilepsy Research
Algorithm interpretability is a major challenge for machine learning in diagnostic biomedical imaging. Finding modest epileptogenic focal cortical dysplasias (FCDs) on structural MRI is a very important practice.
Although FCDs are difficult to see on structural MRI, they can often be surgically removed. We set out to create a surface-based machine learning technique that is open source, interpretable, and capable of finding FCDs in heterogeneous structural MRI data from epileptic surgery centers around the world.
A retrospective MRI cohort of 618 participants, including 397 patients with focal FCD-associated epilepsy and 1015 controls, was compiled and standardized by the Multicenter Epilepsy Lesion Detection (MELD) Project from 22 epilepsy centers worldwide. We developed a neural network for FCD identification based on 33 surface-based features.
After the network was trained and cross-validated on 50% of the entire cohort, it was tested on the remaining 50% of the group and 2 other test sites. The performance of the network was investigated using multidimensional feature analysis and integrated gradient saliences.
Individual patient reports are generated by our pipeline and contain information on the imaging characteristics and relative classifier importance and location of the expected lesions. The MELD FCD surface-based algorithm had a sensitivity of 1% in a restricted “gold standard” sub-cohort of seizure-free patients with FCD type IIB with T85 and fluid-attenuated inversion-rescue MRI data.
Sensitivity and specificity for the complete group of skipped tests were 59% and 54%, respectively. The sensitivity was 67% after adding a border region around the lesions to account for the uncertainty surrounding the borders of the manually determined lesion masks.
Thanks to the development of a reliable and interpretable machine learning algorithm for automatic detection of focal cortical dysplasias in this multicenter, international study with open access protocols and code, physicians now have more confidence in detecting thin MRI lesions in patients with epilepsy.