Selected work

Representative Papers

Can large language models unlock discrete data in ophthalmic diagnostic reports?

Sep 30, 2026

This study addresses the inefficiency and limited accuracy of structured information extraction from ophthalmic diagnostic reports. Leveraging GPT-4o, we compare JSON Schema-constrained generation with pure prompting strategies for automated information extraction across four categories of ophthalmology reports. A Python-based post-processing pipeline is integrated, and performance is evaluated against manual annotation baselines in terms of accuracy and processing time. Results demonstrate that the proposed approach achieves near-perfect accuracy while reducing extraction time by approximately 92% compared to manual annotation. Furthermore, the analysis reveals complementary strengths between the two strategies regarding numerical precision and formatting consistency. Overall, this work validates the efficiency and application potential of general-purpose large language models in processing domain-specific medical documents.

0 citationsRead paper

Longer Records, Broader Invariance: The Hidden Scaling Problem in Longitudinal Contrastive Learning

Sep 28, 2026

This study addresses a critical limitation in longitudinal contrastive learning, where expanding the recording span leads to an excessively broad supervision scope. The resulting over-abundance of positive pairs induces invariance bias, erasing key information about individual temporal dynamics. To overcome this, we formally disentangle the concepts of recording span and supervision span for the first time, proposing a person-level contrastive learning framework that explicitly contrasts other observations from the same individual to reverse information loss. Prospective validation using home sensing data and the GLOBEM cohort demonstrates that our approach effectively mitigates the suppression of change-related signals caused by broad supervision. By significantly recovering individual dynamic features while preserving the advantages of extended recordings, this work establishes a new paradigm for longitudinal representation learning.

0 citationsRead paper

Towards Investigating Residual Hearing Loss: Quantification of Fibrosis in a Novel Cochlear OCT Dataset

Aug 21, 2026

研究使用计算机视觉方法，特别是改进的UNET架构（2D-OCT-UNET），来量化豚鼠耳蜗植入后产生的纤维化情况，以期减少纤维化负担并改善人工耳蜗患者的结果。

0 citationsRead paper

Three-Dimensional Retinal Microvasculature Restoration in OCT Angiography

Jun 03, 2026

This study addresses the challenge of artifacts in optical coherence tomography angiography (OCTA) that severely compromise accurate quantification of retinal blood flow and non-perfused areas. Existing approaches are often limited to two-dimensional processing or fail to fully exploit three-dimensional vascular architecture. To overcome these limitations, this work proposes an end-to-end deep learning algorithm that, for the first time in OCTA, integrates 3D vascular structure by reconstructing a central B-scan from three adjacent input B-scans, thereby preserving spatial resolution while substantially enhancing microvascular fidelity. The model employs an EfficientNet-B5 encoder, a decoder augmented with spatial-channel parallel squeeze-and-excitation modules, and skip connections, trained under supervision using ground truth derived from multi-scan averaging. Experimental results demonstrate significant improvements: PSNR increased to 26.16 ± 1.26 (from 22.23 ± 0.78), SSIM reached 0.91 ± 0.02 (from 0.72 ± 0.03), and 2D and 3D Dice coefficients improved by at least 3.8% and 51.2%, respectively (both p < 0.001).

0 citationsRead paper

Deep Learning-assisted AMD Staging based on OCT and OCT Angiography

Jun 03, 2026

This study addresses the challenge of automated, precise staging diagnosis of age-related macular degeneration (AMD) by systematically comparing the performance of deep learning models based on three distinct input modalities: biomarker maps, 2D en face projections, and 3D OCT/OCTA volumetric data. Utilizing an EfficientNet architecture combined with normalization, data augmentation, and five-fold cross-validation, the study evaluates these inputs on a four-stage AMD classification task. Results demonstrate strong agreement between all models and expert annotations (QWK ≥ 0.83), with the biomarker-based model achieving the highest overall and most balanced performance (QWK = 0.85 ± 0.03) and an F1-score of 0.59 ± 0.14 for early AMD detection. The 2D model excels in identifying non-AMD cases with the highest precision (0.79 ± 0.06). This work provides critical guidance for selecting input strategies in automated AMD staging.

0 citationsRead paper

Recent publications

Latest Papers

Can large language models unlock discrete data in ophthalmic diagnostic reports?

Sep 30, 2026

This study addresses the inefficiency and limited accuracy of structured information extraction from ophthalmic diagnostic reports. Leveraging GPT-4o, we compare JSON Schema-constrained generation with pure prompting strategies for automated information extraction across four categories of ophthalmology reports. A Python-based post-processing pipeline is integrated, and performance is evaluated against manual annotation baselines in terms of accuracy and processing time. Results demonstrate that the proposed approach achieves near-perfect accuracy while reducing extraction time by approximately 92% compared to manual annotation. Furthermore, the analysis reveals complementary strengths between the two strategies regarding numerical precision and formatting consistency. Overall, this work validates the efficiency and application potential of general-purpose large language models in processing domain-specific medical documents.

0 citationsRead paper

Longer Records, Broader Invariance: The Hidden Scaling Problem in Longitudinal Contrastive Learning

Sep 28, 2026

This study addresses a critical limitation in longitudinal contrastive learning, where expanding the recording span leads to an excessively broad supervision scope. The resulting over-abundance of positive pairs induces invariance bias, erasing key information about individual temporal dynamics. To overcome this, we formally disentangle the concepts of recording span and supervision span for the first time, proposing a person-level contrastive learning framework that explicitly contrasts other observations from the same individual to reverse information loss. Prospective validation using home sensing data and the GLOBEM cohort demonstrates that our approach effectively mitigates the suppression of change-related signals caused by broad supervision. By significantly recovering individual dynamic features while preserving the advantages of extended recordings, this work establishes a new paradigm for longitudinal representation learning.

0 citationsRead paper

Towards Investigating Residual Hearing Loss: Quantification of Fibrosis in a Novel Cochlear OCT Dataset

Aug 21, 2026

研究使用计算机视觉方法，特别是改进的UNET架构（2D-OCT-UNET），来量化豚鼠耳蜗植入后产生的纤维化情况，以期减少纤维化负担并改善人工耳蜗患者的结果。

0 citationsRead paper

Three-Dimensional Retinal Microvasculature Restoration in OCT Angiography

Jun 03, 2026

This study addresses the challenge of artifacts in optical coherence tomography angiography (OCTA) that severely compromise accurate quantification of retinal blood flow and non-perfused areas. Existing approaches are often limited to two-dimensional processing or fail to fully exploit three-dimensional vascular architecture. To overcome these limitations, this work proposes an end-to-end deep learning algorithm that, for the first time in OCTA, integrates 3D vascular structure by reconstructing a central B-scan from three adjacent input B-scans, thereby preserving spatial resolution while substantially enhancing microvascular fidelity. The model employs an EfficientNet-B5 encoder, a decoder augmented with spatial-channel parallel squeeze-and-excitation modules, and skip connections, trained under supervision using ground truth derived from multi-scan averaging. Experimental results demonstrate significant improvements: PSNR increased to 26.16 ± 1.26 (from 22.23 ± 0.78), SSIM reached 0.91 ± 0.02 (from 0.72 ± 0.03), and 2D and 3D Dice coefficients improved by at least 3.8% and 51.2%, respectively (both p < 0.001).

0 citationsRead paper

Deep Learning-assisted AMD Staging based on OCT and OCT Angiography

Jun 03, 2026

This study addresses the challenge of automated, precise staging diagnosis of age-related macular degeneration (AMD) by systematically comparing the performance of deep learning models based on three distinct input modalities: biomarker maps, 2D en face projections, and 3D OCT/OCTA volumetric data. Utilizing an EfficientNet architecture combined with normalization, data augmentation, and five-fold cross-validation, the study evaluates these inputs on a four-stage AMD classification task. Results demonstrate strong agreement between all models and expert annotations (QWK ≥ 0.83), with the biomarker-based model achieving the highest overall and most balanced performance (QWK = 0.85 ± 0.03) and an F1-score of 0.59 ± 0.14 for early AMD detection. The 2D model excels in identifying non-AMD cases with the highest precision (0.79 ± 0.06). This work provides critical guidance for selecting input strategies in automated AMD staging.

0 citationsRead paper