Multicenter Computer-Aided Diagnosis for Lymph Nodes Using Unsupervised Domain-Adaptation Networks Based on
RuoXi Qin1, Huike Zhang2, LingYun Jiang1
1PLA Strategy Support Force Information Engineering University, Zhengzhou 450001, China.
Computational and Mathematical Methods in Medicine
|May 27, 2020
Summary
This study introduces a novel domain adaptation method for lymph node CT image analysis. Our approach enhances diagnostic system performance across diverse data sources by preserving crucial class information.
Area of Science:
- Medical Imaging
- Computer-Aided Diagnosis
- Deep Learning
Background:
- Computer-aided diagnosis (CAD) systems for lymph nodes often suffer performance degradation due to data variability from multicenter CT image sources.
- Existing domain adaptation methods struggle with the unique challenges of large CT image sizes and complex data distributions in lymph node analysis.
Purpose of the Study:
- To develop a robust domain adaptation technique for improving the performance of lymph node computer-aided diagnosis systems across multicenter CT data.
- To address the variability adaptation problem specific to lymph node CT images by considering shared features and domain-specific conditioning information.
Main Methods:
- Proposed a novel domain adaptation framework utilizing a cross-domain confounding representation to extract domain-invariant features.
- Implemented a cycle-consistency learning framework to preserve class-conditioning information through cross-domain image translations.
- Employed pixel-level cross-domain image mapping and semantic-level cycle consistency for stable confounding representation.
Main Results:
- The proposed method achieved a significant improvement in accuracy, outperforming existing domain adaptation techniques by at least 4.4% on multicenter lymph node data.
- Demonstrated the effectiveness of the cycle-consistency learning framework in preserving class-conditioning information for enhanced domain adaptation.
- Validated the capability of the approach to handle complex feature distributions and achieve domain invariance.
Conclusions:
- The developed domain adaptation method effectively addresses the challenges of multicenter lymph node CT data, leading to more robust and high-performance computer-aided diagnosis systems.
- The combination of cross-domain confounding representation and cycle-consistency learning offers a promising direction for domain adaptation in medical imaging.
- This approach provides a stable confounding representation with class-conditioning information, crucial for effective adaptation under complex feature distributions.
