Computer Science Faculty Publications
Document Type
Article
Publication Date
9-2026
Abstract
Purpose
To develop a mutual information (MI)–based mechanism for quantifying representation distance, and to introduce a constrained multiview learning paradigm that dynamically re-ranks and selects sample views, thereby improving contrastive representation learning for lung lesion segmentation on CT images—addressing the difficulty of measuring distances in high-dimensional feature spaces and the impracticality of constructing large positive–negative sample banks in the medical domain.Materials and Methods
The proposed framework, termed MIMIC (Mutual Information-based constrained Multi-view Contrastive learning), generates multiple frequency-domain views of CT images and performs a dynamic MI-based representation re-ranking and selection process to improve the quality of positive and negative sample pairs. MIMIC was integrated into two representative segmentation backbones—a CNN-based U-Net and a CNN-Transformer hybrid TransUNet—and evaluated on three public COVID-19 lung CT segmentation. Performance was benchmarked against 16 established CNN- and Transformer-based segmentation models using three metrics: Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and average surface distance (ASD).Results
MIMIC-enhanced backbones achieved the highest DSC among all compared methods on every dataset. On COVID19-CT-100, MIMIC improved U-Net's DSC from 72.1% to 82.3%. On COVID19-CT-Seg20, MIMIC improved TransUNet's DSC from 77.4% to 88.2%. On MosMedData, MIMIC improved U-Net's DSC from 69.2% to 80.3%. MIMIC-enhanced models also generally achieved the lowest HD95 and ASD (e.g., HD95 = 1.65 mm and ASD = 0.15 on COVID19-CT-Seg20, versus 1.16 mm/0.24 for baseline TransUNet), indicating more accurate lesion boundary delineation in addition to higher region overlap. Ablation experiments further confirmed that MI-based view selection and MI-based contrastive learning are individually beneficial but jointly necessary: combining both within MIMIC consistently outperformed either component alone.Conclusions
The proposed MIMIC framework, driven by MI-based view selection and re-ranking, effectively optimizes multiview contrastive representation learning and substantially improves lung lesion segmentation performance across CNN and Transformer backbones. It is model-agnostic and can be integrated into existing segmentation pipelines without architectural modification.Recommended Citation
Dai, Siyuan, Kai Ye, Kun Zhao, Yang Du, Haoteng Tang, and Liang Zhan. "Constrained Multiview Contrastive Learning for Jointly Supervised Representation Learning." Meta-Radiology (2026): 100216. https://doi.org/10.1016/j.metrad.2026.100216
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Publication Title
Meta-Radiology
DOI
10.1016/j.metrad.2026.100216
Comments
© 2026 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. http://creativecommons.org/licenses/by/4.0/