Institution profile

Nanjing Medical University

Academic institutionasia · cn
Official website
Research library15linked papers
Opportunities0open roles
Selected work

Representative Papers

A Lightweight Medical Image Classification Framework via Self-Supervised Contrastive Learning and Quantum-Enhanced Feature Modeling

Jan 23, 2026

This work proposes a lightweight classical-quantum hybrid architecture to address key challenges in medical image analysis, including scarce labeled data, limited computational resources, and poor model generalization. Building upon a MobileNetV2 backbone, the framework leverages SimCLR-style self-supervised contrastive pretraining to learn generalizable representations and integrates a low-parameter parametrized quantum circuit (PQC) module to enhance feature discriminability. To the best of our knowledge, this is the first approach to combine self-supervised learning with lightweight quantum modeling. With only 2–3 million parameters, the model achieves significant performance gains over classical baselines after fine-tuning on small annotated datasets, consistently improving accuracy, AUC, and F1-score. Feature visualizations further confirm the discriminative power and stability of the learned representations.

1 citationsRead paper
Recent publications

Latest Papers

Cross-Modal Ultrasound-MRI Learning for Fetal Brain Ventricular Volumetry and Abnormality Screening

Aug 14, 2026

This study addresses the subjectivity of ultrasound ventricular assessment and limited MRI accessibility by proposing VIFBA, a framework enabling MRI-grade volumetric prediction and abnormality screening using only ultrasound videos. Methodologically, it integrates JEPA-inspired tubular latent space prediction, cross-modal contrastive alignment, and retrieval-augmented vision-language modeling to enhance spatiotemporal feature representation. Validated on 857 paired cases, VIFBA achieved a mean absolute error of 0.59 mL (r = 0.99) for volume regression, 0.94 accuracy in ventriculomegaly classification, and an F1-score of 0.78 for multi-abnormality detection. These results demonstrate that the proposed approach facilitates low-cost, high-precision intelligent assessment of prenatal brain development, effectively bridging the diagnostic gap between conventional ultrasound and neuroimaging standards without requiring expensive hardware or extensive manual annotation.

0 citationsRead paper