DINO-LG: Enhancing Vision Transformers with Label Guidance for Coronary Artery Calcium Detection

summary

The gist

Coronary artery calcium (CAC) scoring via computed tomography (CT) is a key method for prevention in coronary artery disease…

In short

This study introduces DINO-LG, a label-guided extension of Vision Transformers to improve Coronary Artery Calcium (CAC) detection from CT scans. By using targeted augmentation on annotated calcified regions during self-supervised pre-training, the model learns features that concentrate on small calcifications. The resulting three-stage pipeline achieves 90% accuracy in CAC risk classification, outperforming standalone segmentation methods.

Key concepts

DINO
DINO is a popular self-supervised learning technique for Vision Transformers that learns powerful visual representations from unlabeled images without needing manual labels. It works by using self-distillation to train the model to predict its own output, allowing it to capture general image features effectively.
DINO-LG
DINO-LG is a novel method that enhances DINO by incorporating label guidance. It specifically targets annotated calcified regions during pre-training by generating guided crops. This forces the model's self-attention mechanisms to focus on areas of clinical interest, such as calcium deposits.
Three-stage Pipeline
The proposed system uses a three-step process: first, DINO-LG trained Vision Transformer (ViT) extracts features; second, a linear classifier filters slices based on calcification presence; and third, a U-NET model segments the calcified regions for accurate scoring. This integrated approach increases efficiency and accuracy.
Label Guidance
Label guidance is the technique used in DINO-LG to improve feature learning. Instead of relying only on random crops, this method generates specific crops that contain annotated calcified regions alongside random ones. This dual cropping balances learning general image features with capturing the specific characteristics of calcium deposits.

Terminology used across episodes

This episode discusses

The paper

DINO-LG: Enhancing Vision Transformers with Label Guidance for Coronary Artery Calcium Detection · Read on arXiv

University of Kentucky · Department of Software Engineering, Karabuk University

DOI: 10.1007/s11517-026-03523-1

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "DINO-LG: Enhancing Vision Transformers with Label Guidance for Coronary Artery Calcium Detection".

Dev: Here's the gist of the paper. Coronary artery calcium (CAC) scoring via computed tomography (CT) is a key method for prevention in coronary artery disease, and this study introduces DINO-LG to enhance Vision Transformers with label guidance to improve CAC detection accuracy Pruning redundant reasoning tokens DINO-LG is a novel label-guided extension of DINO that incorporates targeted augmentation on annotated calcified regions during self-supervised pre-training The three-stage pipeline integrates Vision Transformer (ViT-Base/8) feature extraction via DINO-LG trained on 914 CT scans comprising 700 gated and 214 non-gated acquisitions, linear classification to identify calcified slices, and U-NET segmentation for CAC quantification and Agatston scoring This integrated system achieves 90% accuracy in CAC risk classification on 45 test patients, outperforming standalone U-NET segmentation (76% accuracy) while processing only the relevant subset of CT slices How the training signal is built The label-guided process begins by selecting random coordinates from within the annotated calcified areas to generate focal points for generating crops that contain the regions of interest This dual cropping approach enables the model to learn both calcification-specific features and general image representations In our implementation, we generate 4 guided crops and 8 random crops per labeled image to balance specific feature learning with general representation capabilities The figure shows that DINO-LG features demonstrate concentrated activation on calcified regions, validating successful feature guidance toward regions of clinical interest Where the gains come from The classification stage leverages the discriminative power of DINO features to effectively distinguish between slices with and without calcification The classifier is trained using a balanced loss function that accounts for the class imbalance, ensuring that both positive (calcified) and negative (non-calcified) slices are correctly classified The reduction in false-negative predictions is clinically significant as these represent missed diagnoses where calcified slices are incorrectly classified as non-calcified Performance analysis of the integrated system The integrated system demonstrates significant improvement in identifying and scoring coronary artery calcification In the classification part, the DINO-LG model efficiently distinguishes CT slices whether having calcified areas or not, ensuring that only relevant images are forwarded to the segmentation model The segmentation model used in the system is U-NET, chosen for its superior performance in IoU and Dice scores The results show improved performance of the integrated system in LAD and LCX regions, where smaller calcifications are more prevalent The overall weighted averages for precision, recall, and F1-score in our approach are 0.90 each, substantially outperforming the corresponding values of 0.80, 0.76, and 0.73 in the U-NET model Clinical implications and limitations The high specificity (90%) of our classification stage reduces unnecessary segmentation processing, which is important for clinical workflow integration The ability to process both ECG-gated and non-gated CT acquisitions with a single model could facilitate opportunistic CAC screening from routine chest CT examinations However, the scale of our training dataset (1,003 patients total) is relatively small for foundational model development The study focuses on technical performance metrics without correlation to long-term cardiovascular outcomes, which is essential for demonstrating clinical utility Detection of very small calcifications (< 1mm) remains challenging, particularly in cases with high image noise Additional limitations include the computational overhead of the three-stage pipeline and lack of robustness testing against common CT artifacts like motion blur or metal artifacts The ability of vision foundational models to capture multiple feature patterns simultaneously suggests potential for cross-domain learning across different anatomical regions Comparison with prior work Compared to recent deep learning approaches for CAC scoring, our method offers several advantages While Eng et al.

32: achieved 71-94% sensitivity using supervised learning on dedicated datasets, our self-supervised approach requires fewer annotated samples while maintaining competitive performance (89% sensitivity) The classification-then-segmentation pipeline also provides computational advantages by processing only relevant slices, reducing inference time by approximately 60% compared to volume-wide segmentation approaches Our unified framework handles both ECG-gated and non-gated CT acquisitions, addressing a practical clinical need that most prior work has not explored systematically The superior performance of DINO-LG stems from its ability to learn calcification-specific features through targeted augmentation Analysis of the attention maps reveals that label-guided training directs the model’s self-attention mechanisms toward high-density regions corresponding to calcifications The PCA visualization demonstrates that label-guided training produces features with concentrated activation on calcified regions, contrasting sharply with the diffuse patterns observed in standard DINO features The classification stage reduced false-negative predictions by 49% and false-positive predictions by 57% This enhanced performance is particularly beneficial in clinical settings where accurate CAC assessment directly impacts patient management decisions This work demonstrates that label-guided self-supervised learning can effectively direct Vision Transformers to learn task-specific features while maintaining the benefits of pre-training on unlabeled data DINO-LG achieves substantial improvements in coronary artery calcium detection and scoring, offering a promising approach for medical imaging tasks where labeled data is scarce and regions of interest are small and sparse The integration of self-supervised learning, targeted classification, and segmentation provides both improved accuracy and computational efficiency The ability to process both gated and non-gated scans with a single model could facilitate opportunistic CAC screening from routine chest CT examinations potentially identifying at-risk patients who would not otherwise undergo dedicated cardiac imaging The study shows that the integrated system achieves 90% accuracy in CAC risk classification compared to 76% for standalone segmentation The superior performance of DINO-LG stems from its ability to learn calcification-specific features through targeted augmentation Analysis of the attention maps reveals that label-guided training directs the model’s self-attention mechanisms toward high-density regions corresponding to calcifications The PCA visualization demonstrates that label-guided training produces features with concentrated activation on calcified regions, contrasting sharply with the diffuse patterns observed in standard DINO features The classification stage reduced false-negative predictions by 49% and false-positive predictions by 57% This enhanced performance is particularly beneficial in clinical settings where accurate CAC assessment directly impacts patient management decisions The system's versatility across different imaging protocols and strong alignment with expert evaluations demonstrate its potential for broad clinical use Future work should focus on further reducing the 11% false-negative rate in slice classification through ensemble methods or uncertainty quantification The study shows that label-guided training produces features with concentrated activation on calcified regions, contrasting sharply with the diffuse patterns observed in standard DINO features The integrated system achieves consistent improvements across all risk categories, particularly excelling in precision, recall, and F1-scores for both low-risk and high-risk categories The average sensitivity (TPR) across all risk categories rises from 0.69 in the U-NET model to 0.86 in the proposed system The overall weighted averages for precision, recall, and F1-score in our approach are 0.90 each, substantially outperforming the corresponding values of 0.80, 0.76, and 0.73 in the U-NET model The U-NET architecture achieves the highest average performance across all coronary artery classes with IoU scores of 1.00 for RCA The variation in performance across different arteries reflects the inherent challenges of detecting smaller calcifications in LAD and LCX compared to the typically larger and higher-contrast calcifications in RCA The integrated system partially addresses this limitation, achieving improved F1-scores of 0.85 for LAD and 0.80 for LCX compared to 0.76 and 0.75 for standalone U-NET The training method we propose for DINO is not limited to coronary artery calcium detection; this approach can be applied to various medical imaging tasks with sparse annotations including lung nodule detection, liver lesion identification, and tumor segmentation Page 1 --- DINO-LG: Enhancing Vision Transformers with Label Guidance for Coronary Artery Calcium Detection Mahmut S.

Dev: Gokmen1, Caner Ozcan1,2, Moneera N.

Rosa: Haque1, Steve W.

Title and authors: Rosa: So we're talking about this paper called "DINO-LG: Enhancing Vision Transformers with Label Guidance for Coronary Artery Calcium Detection." It sounds pretty technical, but basically, they’re trying to make the AI much better at finding calcium in chest CT scans.

Dev: Yeah, it's focused on coronary artery disease risk assessment using computed tomography or CT scans. The authors are Mahmut Gokmen and the team from University of Kentucky and Karabuk University. They're taking a foundational model called DINO, which is a self-distillation model without any labels, and they’re extending it with label guidance to help it see calcified regions specifically.

Rosa: So the title tells us they are using this label guidance to enhance vision transformers for CAC detection. It suggests that standard vision transformers might struggle when the target object, which is tiny calcium deposits, is so sparse in the image.

Dev: Exactly. They’re trying to fix that scarcity problem by giving the model some explicit direction on where to look during its training process. It’s about improving diagnostic precision while lowering healthcare costs by minimizing unnecessary tests and treatments <ref:2411.07976#pg1>

Taro: From my side, I'm thinking about how this relates to real-world failure modes. If the AI is struggling with those small calcifications, what happens when it encounters a noisy or low-contrast scan in a clinical setting? Does this guidance help it stay focused when the input data is messy?

Rosa: That’s a good question. The paper mentions that they used nine hundred fourteen CT scans to train their DINO-LG model, which included both gated and non-gated acquisitions. It also focuses on the idea of using both types of scans together in a unified framework <ref:2411.07976#pg2>

The paper's summary: Dev: So the core idea is this three-stage pipeline they propose: first, you extract features using the DINO-LG trained vision transformer. Second, you use a linear classifier to sort through all the CT slices and only keep the ones that actually have calcification. And third, a U-NET segmentation model does its job on those relevant slices to quantify the calcium and calculate an Agatston score.

Rosa: It’s smart because it breaks down the task. Instead of one big model trying to do everything, you have specialized steps. This classification stage is key because it filters out most of the irrelevant images before you spend time on the heavy segmentation part <ref:2411.07976#pg1>

Taro: I wonder about that filtering step. If that classifier misses a slice with calcification, how does the whole system handle that? Does it just miss a diagnosis, or is there a mechanism to recover?

Dev: The paper discusses this in terms of false negatives and false positives. They found that their DINO-LG model achieved eighty-nine percent sensitivity and ninety percent specificity for detecting those calcium-containing CT slices compared to standard DINO which got seventy-nine percent sensitivity and seventy-seven percent specificity <ref:2411.07976#pg2>

Rosa: That difference in sensitivity is substantial when you think about missing a diagnosis that could lead to aggressive treatment decisions based on the calcium score. They also noted they reduced false negative predictions by forty-nine percent and false positive predictions by fifty-seven percent <ref:2411.07976#pg2>

The paper's improvements: Rosa: The paper highlights a specific improvement in how they build the training signal. They use a dual cropping approach for labeled images, generating four guided crops and eight random crops per labeled image. This helps the model learn both the specific calcification features and general image representations simultaneously <ref:2411.07976#pg9>

Dev: That’s where they address those small calcifications you mentioned earlier. The paper shows that these DINO-LG features have concentrated activation on the calcified regions, which is a big deal because standard DINO features were diffuse across cardiac structures <ref:2411.07976#pg10>

Taro: So the label guidance isn't just noise; it’s actively steering the model’s attention mechanism to where it needs to be. When you look at those attention maps, what does that actually look like in practice?

Rosa: The paper says the label-guided training directs the self-attention mechanisms toward high density regions corresponding to calcifications <ref:2411.07976#pg10> It’s making sure the model pays more attention to the right parts of the image.

Dev: And this leads directly into how they improve performance on specific areas. For instance, they saw improved F1 scores of zero point eight five for low-risk and zero point eight zero for medium risk calcium regions in LAD and LCX compared to what a standalone U-NET model achieved <ref:2411.07976#pg14>

Rosa: It seems the whole point is that this integrated system achieves ninety percent accuracy in CAC risk classification on forty five test patients, which is significantly better than the seventy-six percent accuracy of just doing standalone U-NET segmentation <ref:2411.07976#pg2>

Conclusion: Dev: So to wrap up, DINO-LG shows that label-guided self-supervised learning can effectively direct vision transformers to learn task specific features while still benefiting from pre training on unlabeled data. It’s a method that works well when labeled data for small and sparse regions is scarce <ref:2411.07976#pg1>

Rosa: The integration of this classification step, the feature extraction via DINO-LG, and the U-NET segmentation provides both better accuracy and computational efficiency. Plus, it lets them process both ECG-gated and non-gated CT scans with just one model <ref:2411.07976#pg3>

Taro: I still have a question about those limitations. The authors flag that their training dataset size, which is one thousand three patients total, is relatively small for developing a foundational model. Also, they admit that detecting very small calcifications, less than one millimeter in size, remains challenging especially when there's high image noise <ref:2411.07976#pg1>

Dev: That’s fair. The study focuses on technical performance metrics without linking them to long-term cardiovascular outcomes, which is something that’s essential for showing real clinical utility <ref:2411.07976#pg5>

Rosa: Yeah, so the big implication is that this approach could facilitate opportunistic CAC screening from routine chest CT scans. It could potentially identify at-risk patients who wouldn't otherwise get dedicated cardiac imaging <ref:2411.07976#pg3>

Taro: I just want to push on that small calcification limitation. If the method struggles with features smaller than one millimeter, it might not be reliable enough for routine screening of those very small lesions <ref:2411.07976#pg1>

More episodes

← Home