AI Models Evaluated for Spotting Ischemic Stroke Lesions on Routine MRI

Many centers first acquire contrast-free MRI FLAIR sequences because they are quick and broadly available in routine care. These images are challenging for automated analysis due to faint, irregular lesions and scanner variability. A new study published ahead of print in the Journal of Intelligent Systems in Current Computer Engineering evaluates three deep learning methods to improve lesion identification on FLAIR.

The evaluation compared three deep-learning approaches: a conventional 3D U-Net, the self-configuring nnU-Net framework, and an adapted YOLOv8 model configured for 2D segmentation. The researchers aimed to determine which method could most accurately delineate stroke lesions on contrast-free MRI under conditions reflecting routine clinical imaging. To enable a direct comparison, all three models were implemented within the same training and preprocessing pipeline.

The models were trained and tested using the ISLES 2022 benchmark, which includes 250 stroke MRI cases. Identical preprocessing and evaluation procedures were applied across all methods to minimize methodological differences. Segmentation performance was assessed using the Dice similarity coefficient and Intersection over Union, established measures of overlap between model-generated segmentations and expert annotations.

The self-configuring nnU-Net achieved the highest average agreement with expert-marked lesions, with a Dice score of approximately 0.483. The adapted YOLOv8 model performed similarly, with a Dice score of about 0.476, while processing images substantially faster than the other approaches, suggesting potential advantages when throughput is important. The manually tuned 3D U-Net performed less well overall, achieving a Dice score of approximately 0.294 and highlighting its greater sensitivity to parameter selection and variability in multicenter MRI data.

Error analysis showed that all three approaches struggled with small peripheral lesions, diffuse low‑contrast injury, and cases with imperfect spatial alignment. Preprocessing steps such as harmonizing voxel spacing, normalizing intensity, and training on smaller image patches improved model behavior. The authors recommend prioritizing self‑configuring frameworks like nnU‑Net for robustness when compute permits and considering lightweight architectures like YOLOv8 when speed is critical. They emphasize that rigorous multi‑site validation across scanners and institutions is essential before deployment in clinical workflows.