(주) 뷰노

(주) 뷰노

Submit Inquiries
KR EN
VUNO MED INC.
Publications
BACK TO LIST
ISMRM 2025

Discriminative Feature Learning for Lacune Detection in 2D T2-FLAIR Images using Supervised Contrastive Learning

  • May. 2025

Introduction
Cerebral small vessel disease (cSVD) is a progressive neurological condition that impacts brain health through damage to small cerebral blood vessels, predominantly affecting white matter regions. This condition disproportionately affects the elderly population, leading to cognitive decline and elevated stroke risk. The accurate identification of lacunar infarcts has become increasingly crucial for both diagnosing cSVD and determining eligibility for emerging anti-amyloid therapies, as recent studies suggest that the presence of lacunar infarcts may influence treatment response and safety profiles in patients receiving these novel treatments. While brain MRI remains the primary diagnostic tool,
current practice relies heavily on time-intensive radiological interpretation, which is susceptible to inter-reader variability. Recent deep learning approaches for automation face two critical challenges: differentiating lacunes from lacune-mimicking structures (particularly perivascular spaces and vessels) and handling severe data imbalance where normal cases significantly outnumber pathological ones4. To address these challenges, we propose a supervised contrastive learning approach designed to enhance feature discrimination capabilities within imbalanced datasets. This study aimed to develop an automated method that can effectively differentiate lacunes from similar-appearing structures in imbalanced
datasets.


Methods and Materials

  • Data: A total of 427 T2-FLAIR MRI scans (285 positive, 142 negative cases, slice thickness: 4mm, no gap) from the Asan Medical Center were collected, with lacunar infarcts labelled by one experienced radiologists (14 years of experience in neuroradiology). The dataset included subjects with mean age of 73.5 ± 7.8 years (185 males, 242 females) and was split into train/test sets (8:2 ratio). We assigned 20% of the training data as a validation set, maintaining balanced demographic and lesion pattern distributions. The test set contained 54 positive cases (160 lacunar infarcts) and 32 negative cases. Detailed dataset characteristics and distributions can be found in Figure 1. 
  • Pre-processing: All MRIs underwent preprocessing using VUNO Med-DeepBrain software for both intracranial volume (ICV) segmentation5 and white matter hyperintensities (WMH) segmentation (Figure 2., validation DSC: 0.988, 0.855, respectively). Since lacunar infarcts are frequently associated with WMH regions7, the segmentation results were used to select slices containing WMH, effectively reducing the search space while maintaining detection sensitivity. 
  • Method: The framework comprises two stages: 1) representation learning; and 2) segmentation with fine-tuning. Using ResNet-348 as an encoder, we adopted supervised contrastive learning9 to enhance feature discrimination (Figure 3.a). Patches were randomly extracted from FLAIR images and classified based on lacune segmentation masks: patches containing lacunes were treated as positive samples, while patches without lacunes (containing normal tissue, perivascular spaces, or vessels) were treated as negative samples. This approach maximizes the similarity between positive samples while separating them from negative samples in the embedding space, enabling better discrimination of lacunes from similar-appearing structures. The pre-trained encoder was then integrated into the Attention U-Net architecture10 for lacune segmentation (Figure 3.b, learning rate: 1e-5, batch size: 16, epochs: 100). Performance was evaluated using instance-level metrics (Area Under the ROC Curve [AUC], average precision (mean precision across all recall values), sensitivity and specificity), with true positives defined as predictions within 3mm of ground truth lesion centers.


Results
We compared three Attention U-Net variants with different encoder initializations: random
initialization, BraTS11,12,13 pre-trained ResNet-34, and our proposed Supervised Contrastive Learning (SupCon) pre-trained ResNet-34 (Table 1). The random initialized model achieved an AUC of 0.732 and average precision of 0.241, while the BraTS11,12,13 pre-trained encoder improved performance with an AUC of 0.747 and average precision of 0.376. The SupCon encoder showed the best performance across all metrics (AUC: 0.804, average precision: 0.504). As shown in Figure 4, GradCAM visualization revealed that SupCon pre-training enables more precise localization of lacunes compared to other methods, demonstrating the importance of encoder pre-training for detecting small, imbalanced lesions in cSVD.


Discussion
In detecting small lesions with severe data imbalance, such as in cSVD, the encoder's discriminative capability becomes crucial. Our performance improvements demonstrate that supervised contrastive learning can effectively enhance the encoder's ability to distinguish subtle features, particularly in separating lacunes from similar-appearing structures. However, important technical limitations remain. The thick-slice nature of 2D FLAIR images (4mm) impedes detection of subtle lacunes, particularly those partially captured between slices, affecting recall performance. Additionally, the severe class imbalance leads to a precision-recall trade-off favoring false positive reduction over complete detection.


Conclusion
This study presents an effective strategy for training encoders to detect small lesions with similar features in cSVD, through supervised contrastive learning. Future work will focus on expanding the current approach to include lacune localization within specific brain regions and external validation across multiple centers.

Link

-

Author

Saehyun Kim, M.S., Chong Hyun Suh, M.D., Ph.D., Min Woo Han, Wooseok Jung, M. Math., Seung Hyun Lee, M.D.

Tag in

#medical_image

#VUNO Med®-DeepBrain®

Previous
No previous post
Next
Diagnosis of Frontotemporal Dementia on Brain MR Images by Using Automated Brain Volumetry 05. 15. 2025
List

We use cookies or similar technologies as specified in our privacy policy to enhance your experience.

If you would like to learn more about how we use cookies, click "Privacy Policy".