News Release

Detection of vascular pathologies in non-contrast magnetic resonance images for early screening

Peer-Reviewed Publication

Bentham Science Publishers

The study, “Detection of Vascular Pathologies in Non-contrast Magnetic Resonance Images for Early Screening,” was carried out by Patricia García-Berlanga, Juan Zapata, Juan Martínez-Alajarin, and Pablo Hernandez-Cerdan in the Journal of Intelligent Systems in Current Computer Engineering.

Spanish researchers have compared three types of artificial intelligence software on a practical question: can a computer find stroke damage in routine brain MRI scans taken without contrast dye? The self-configuring nnU-Net system performed best, and a lightweight YOLOv8 model came very close. The authors say the overall accuracy is still moderate, so the results are an early step and not a tool ready for hospitals.




If you want to publish your article please visit : https://bit.ly/4de0DRi

Finding Stroke Damage Without Contrast Dye

Ischemic stroke is a leading cause of death and long-term disability worldwide, and deciding quickly how to treat a patient depends on locating the damaged brain tissue. FLAIR is a standard MRI sequence that shows swelling and old strokes without an injected contrast agent, and it is available on most scanners. That makes it attractive for large-scale or automated screening. It is also hard for software to read. Stroke lesions can be small, irregular, and faint, and they can look like harmless age-related changes in the brain’s white matter or like image artefacts. Scanner differences add further variation. Deep learning has improved medical image analysis, but few studies have compared fundamentally different approaches under identical conditions on FLAIR images alone. This study sets out to do that.


Three Approaches, One Test

The team used the public ISLES 2022 dataset of 250 stroke cases with expert-drawn lesion outlines, and trained all three models on FLAIR scans only, with the same data split and the same preprocessing wherever possible. The first model was a manually configured 3D U-Net, a long-standing standard in medical imaging. The second was nnU-Net, which adjusts its own settings to suit the data. The third was YOLOv8, a fast model built for everyday object detection, adapted here to work through the scan slice by slice and then rebuilt into a 3D result. Results were judged on 25 test cases using the Dice score, where 1 means a perfect match with the expert outline. nnU-Net scored 0.483 and YOLOv8 scored 0.476, with the 3D U-Net well behind at 0.294. nnU-Net improved the average Dice score by about 64 percent over the U-Net and produced smoother, more coherent lesion outlines, including for oddly shaped lesions. YOLOv8 matched it closely in many cases but tended to miss small, diffuse, or faint lesions. All three models varied widely from case to case, and nnU-Net still returned no lesion at all for a few scans.


What the Findings Mean

The authors conclude that nnU-Net is the better choice when accuracy and consistency matter most and computing power allows. YOLOv8 is a practical option where simplicity and speed come first, such as quick screening. The classic U-Net remains a useful reference point for comparison. They are careful about the limits of the work. Only 25 cases were used for testing, only one training run (fold) of nnU-Net was trained because of limited computing resources, and the results have not been checked on scans from other hospitals. YOLOv8’s speed advantage follows from its design but was not formally measured. The study also used only FLAIR, without the other MRI sequences in the dataset that might help. For these reasons the authors describe the results as exploratory. They recommend testing on larger and more varied datasets, combining several MRI sequences, and exploring hybrid systems that pair the precision of 3D models with the speed of 2D ones. If it proves reliable, contrast-free automated screening could help radiologists spot subtle stroke damage, support more consistent readings, and reduce the need for contrast in patients who cannot receive it. The research was led by corresponding author Patricia García-Berlanga of the Electronics, Computer Technology and Projects Department, Universidad Politécnica de Cartagena, Member of European University of Technology EUT+, Cartagena, Spain, with co-authors Juan Zapata and Juan Martínez-Alajarin from the same department, and Pablo Hernandez-Cerdan of the I+D Department at Vasomaly, Torre-Pacheco, Murcia, Spain. The work was supported by Spain’s Ministry of Science, Innovation and Universities (grant PID2021-128009OBC33).

Read the published article here: https://bit.ly/4zCchiD
 


Article title: Detection of Vascular Pathologies in Non-contrast Magnetic Resonance Images for Early Screening

DOI: 10.2174/0130505070466701260824140841
http://dx.doi.org/10.2174/0130505070466701260824140841


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.