Supervised Contrastive Learning for Lacune Detection in MRI
Original Title: Biomarkers
Journal: Alzheimer's & dementia : the journal of the Alzheimer's Association
Overview
Lacunes are small, deep brain infarcts that indicate vascular disease and increase the risk of cognitive decline. Detecting these features manually is time-consuming and prone to error due to their small size and similarity to other structures like perivascular spaces. This study presents a deep learning framework designed to automate the segmentation of lacunes using 2D T2-FLAIR MRI scans. The researchers utilized a dataset of 427 images, which underwent preprocessing to segment intracranial volume and white matter hyperintensities. The core architecture employed is an Attention U-Net. To address the challenge of imbalanced data where lacunes are rare compared to healthy tissue, the model incorporates a ResNet-34 encoder pretrained with supervised contrastive learning. This approach allows the model to learn discriminative features by grouping similar positive instances and separating them from negative ones. Evaluation on a test set showed the model identified 102 out of 166 lacunes (61.5%), yielding a figure-of-merit of 0.726. At the patient level, the system achieved an area under the curve of 0.810 for detecting the presence of lacunes.
Novelty
The primary methodological advancement lies in the integration of supervised contrastive learning within a segmentation pipeline for rare radiological features. Traditional segmentation models often struggle with class imbalance, frequently misidentifying lacune-mimicking structures as true infarcts. By employing supervised contrastive learning during the pretraining phase, the encoder learns to map lacune samples closer together while pushing them away from non-lacune samples, such as prominent vessels or white matter lesions. This contrastive approach enhances the model's ability to extract semantic features before the actual segmentation task begins. Furthermore, the use of an Attention U-Net architecture helps the model focus on relevant spatial regions, which is critical given the minute scale of lacunar infarcts relative to the entire brain volume. This combination of contrastive representation learning and spatial attention specifically targets the difficulty of distinguishing small, infrequent pathologies from common anatomical variants.
Potential Clinical / Research Applications
The model provides a foundation for automated vascular health assessment in aging populations. In clinical practice, it could serve as a second-reader tool to assist radiologists in identifying subtle lacunes that might otherwise be overlooked during routine screenings for dementia or stroke risk. Because lacunes are linked to an increased risk of amyloid-related imaging abnormalities, this tool could be valuable in monitoring patients undergoing monoclonal antibody therapies for Alzheimer's disease. In a research context, the automated quantification of lacunar burden allows for large-scale longitudinal studies investigating the progression of small vessel disease and its correlation with cognitive impairment. Future refinements could include localizing lacunes within specific functional brain regions, such as the thalamus or basal ganglia, to predict specific neurological deficits. This would enable more personalized risk profiling and management strategies for patients at risk of vascular-related cognitive decline.