Selected work

Representative Papers

Clinically-aligned ischemic stroke segmentation and ASPECTS scoring on NCCT imaging using a slice-gated loss on foundation representations

Feb 27, 2026

This study addresses a critical limitation in current deep learning approaches for segmenting ischemic stroke lesions on non-contrast CT (NCCT) scans: the neglect of the anatomical coupling between the basal ganglia (BG) and supraganglionic (SG) levels as defined by the ASPECTS scoring system, which undermines alignment with clinical evaluation logic. To bridge this gap, the authors propose the first integration of this structured clinical prior into foundation model training through an anatomy-aware gated loss (TAGL). By combining a frozen DINOv2 backbone with a lightweight decoder, their method enforces BG-SG consistency without increasing inference overhead. Evaluated on the AISD dataset, the approach achieves a Dice score of 0.6385, outperforming existing CNNs and foundation models. On an in-house ASPECTS dataset, it further improves the average Dice from 0.698 to 0.767, significantly enhancing segmentation consistency with clinical standards.

0 citationsRead paper

Extending 2D foundational DINOv3 representations to 3D segmentation of neonatal brain MR images

Feb 27, 2026

This work addresses the challenge of directly applying 2D foundation models to 3D hippocampal segmentation in neonatal brain MRI by proposing a structured window decomposition–recomposition strategy. The method partitions the 3D volume into non-overlapping sub-cubes, leverages frozen 2D DINOv3 features for parallel decoding, and reconstructs a full 3D segmentation via a dense prediction head. This approach maintains constant decoder memory usage while effectively recovering 3D anatomical structures from frozen 2D representations, enabling generalizable extension to 3D medical imaging. Evaluated on the ALBERT dataset, the method achieves a Dice score of 0.65 for single-window hippocampal segmentation, demonstrating both its efficacy and anatomical consistency.

0 citationsRead paper

Rethinking Test-Time Training: Tilting The Latent Distribution For Few-Shot Source-Free Adaptation

Feb 02, 2026

This work proposes a training-free inference method for few-shot test-time adaptation under the challenging setting where the model is entirely frozen and source data are inaccessible. The approach performs a measure transformation on the encoder’s latent embedding distribution via exponential tilting, leveraging the support set to compute task similarity scores and reweighting the distribution according to a KL-optimal criterion to adapt to the new task. As the first method to achieve training-free test-time adaptation with a completely frozen model, it relies solely on inference-stage distribution correction. Extensive experiments demonstrate that the proposed method matches or even surpasses the performance of approaches requiring parameter updates across multiple few-shot benchmarks, thereby validating the effectiveness and potential of purely inference-level adaptation.

0 citationsRead paper

Multi-Task Learning for Visually Grounded Reasoning in Gastrointestinal VQA

Nov 06, 2025

To address weak interpretability and insufficient reasoning capability in gastrointestinal endoscopic visual question answering (VQA), this paper proposes a LoRA-based multi-task collaborative learning framework that jointly models VQA, natural language explanation generation, and visual grounding. Methodologically, we adopt Florence-2 as the backbone and integrate Kvasir-VQA-x1, synthetically generated explanations, and text-region alignment data to achieve cross-modal semantic alignment and joint modeling of medical logic. Our key contribution is the first end-to-end, multi-task interpretable reasoning architecture for gastrointestinal medical VQA, enabled by parameter-efficient fine-tuning to ensure task synergy and generalization. Experiments demonstrate significant improvements over single-task baselines: +4.2% in answer accuracy and +6.8% in grounding IoU—validating the effectiveness of multi-task learning in enhancing both medical visual reasoning and model interpretability.

0 citationsRead paper

Comparative Study of CNN Architectures for Binary Classification of Horses and Motorcycles in the VOC 2008 Dataset

Nov 06, 2025

This study addresses the severe class imbalance in the horse vs. motorcycle binary classification task on PASCAL VOC 2008. We systematically evaluate nine state-of-the-art CNN architectures—including ConvNeXt-Tiny, ResNet-50, DenseNet-121, and ViT—under minority-class–specific data augmentation. Employing a unified training protocol and rigorous evaluation, we find ConvNeXt-Tiny achieves the highest mean average precision (mAP): 95.53% for horses and 89.12% for motorcycles. Empirical results demonstrate that targeted augmentation significantly improves minority-class detection, particularly boosting motorcycle recall. To our knowledge, this is the first work to quantitatively benchmark the robustness of diverse modern architectures to class imbalance in a fine-grained binary object classification setting. Our reproducible experimental framework establishes a practical, lightweight, and high-accuracy baseline for imbalanced visual classification, offering actionable insights for architecture selection and data-augmentation strategy design.

0 citationsRead paper

Recent publications

Latest Papers

Clinically-aligned ischemic stroke segmentation and ASPECTS scoring on NCCT imaging using a slice-gated loss on foundation representations

Feb 27, 2026

This study addresses a critical limitation in current deep learning approaches for segmenting ischemic stroke lesions on non-contrast CT (NCCT) scans: the neglect of the anatomical coupling between the basal ganglia (BG) and supraganglionic (SG) levels as defined by the ASPECTS scoring system, which undermines alignment with clinical evaluation logic. To bridge this gap, the authors propose the first integration of this structured clinical prior into foundation model training through an anatomy-aware gated loss (TAGL). By combining a frozen DINOv2 backbone with a lightweight decoder, their method enforces BG-SG consistency without increasing inference overhead. Evaluated on the AISD dataset, the approach achieves a Dice score of 0.6385, outperforming existing CNNs and foundation models. On an in-house ASPECTS dataset, it further improves the average Dice from 0.698 to 0.767, significantly enhancing segmentation consistency with clinical standards.

0 citationsRead paper

Extending 2D foundational DINOv3 representations to 3D segmentation of neonatal brain MR images

Feb 27, 2026

This work addresses the challenge of directly applying 2D foundation models to 3D hippocampal segmentation in neonatal brain MRI by proposing a structured window decomposition–recomposition strategy. The method partitions the 3D volume into non-overlapping sub-cubes, leverages frozen 2D DINOv3 features for parallel decoding, and reconstructs a full 3D segmentation via a dense prediction head. This approach maintains constant decoder memory usage while effectively recovering 3D anatomical structures from frozen 2D representations, enabling generalizable extension to 3D medical imaging. Evaluated on the ALBERT dataset, the method achieves a Dice score of 0.65 for single-window hippocampal segmentation, demonstrating both its efficacy and anatomical consistency.

0 citationsRead paper

Rethinking Test-Time Training: Tilting The Latent Distribution For Few-Shot Source-Free Adaptation

Feb 02, 2026

This work proposes a training-free inference method for few-shot test-time adaptation under the challenging setting where the model is entirely frozen and source data are inaccessible. The approach performs a measure transformation on the encoder’s latent embedding distribution via exponential tilting, leveraging the support set to compute task similarity scores and reweighting the distribution according to a KL-optimal criterion to adapt to the new task. As the first method to achieve training-free test-time adaptation with a completely frozen model, it relies solely on inference-stage distribution correction. Extensive experiments demonstrate that the proposed method matches or even surpasses the performance of approaches requiring parameter updates across multiple few-shot benchmarks, thereby validating the effectiveness and potential of purely inference-level adaptation.

0 citationsRead paper

Multi-Task Learning for Visually Grounded Reasoning in Gastrointestinal VQA

Nov 06, 2025

To address weak interpretability and insufficient reasoning capability in gastrointestinal endoscopic visual question answering (VQA), this paper proposes a LoRA-based multi-task collaborative learning framework that jointly models VQA, natural language explanation generation, and visual grounding. Methodologically, we adopt Florence-2 as the backbone and integrate Kvasir-VQA-x1, synthetically generated explanations, and text-region alignment data to achieve cross-modal semantic alignment and joint modeling of medical logic. Our key contribution is the first end-to-end, multi-task interpretable reasoning architecture for gastrointestinal medical VQA, enabled by parameter-efficient fine-tuning to ensure task synergy and generalization. Experiments demonstrate significant improvements over single-task baselines: +4.2% in answer accuracy and +6.8% in grounding IoU—validating the effectiveness of multi-task learning in enhancing both medical visual reasoning and model interpretability.

0 citationsRead paper

Comparative Study of CNN Architectures for Binary Classification of Horses and Motorcycles in the VOC 2008 Dataset

Nov 06, 2025

This study addresses the severe class imbalance in the horse vs. motorcycle binary classification task on PASCAL VOC 2008. We systematically evaluate nine state-of-the-art CNN architectures—including ConvNeXt-Tiny, ResNet-50, DenseNet-121, and ViT—under minority-class–specific data augmentation. Employing a unified training protocol and rigorous evaluation, we find ConvNeXt-Tiny achieves the highest mean average precision (mAP): 95.53% for horses and 89.12% for motorcycles. Empirical results demonstrate that targeted augmentation significantly improves minority-class detection, particularly boosting motorcycle recall. To our knowledge, this is the first work to quantitatively benchmark the robustness of diverse modern architectures to class imbalance in a fine-grained binary object classification setting. Our reproducible experimental framework establishes a practical, lightweight, and high-accuracy baseline for imbalanced visual classification, offering actionable insights for architecture selection and data-augmentation strategy design.

0 citationsRead paper