TL;DR
This paper introduces FCITAFR, a novel framework for incomplete multiview clustering (IMVC) that tackles the challenge of missing data in medical imaging, specifically for postoperative liver tumor diagnosis. By reconstructing existing features before predicting missing ones and integrating fuzzy clustering with information theory, FCITAFR achieves superior clustering performance compared to existing methods, even with significant data loss.
What Problem Does This Paper Address?
Multiview clustering (MVC) is powerful for leveraging diverse data sources, but real-world scenarios, especially in healthcare, often suffer from incomplete data. This is particularly true for medical imaging like multiphase contrast-enhanced computed tomography (CECT) scans used for liver tumor diagnosis. Patient privacy concerns, data transmission issues, or simply incomplete scans mean that not all views (e.g., non-contrast, arterial phase, portal venous, delay phase) are available for every patient. This "incomplete multiview clustering" (IMVC) problem is challenging because existing methods often struggle with:
- Reliance on paired views: Many methods assume that if one view is present, its corresponding paired view is also available, which is often not the case in clinical practice.
- Direct prediction pitfalls: Simply predicting missing features can lead to errors if key discriminative information is overlooked, negatively impacting clustering accuracy.
- Subtle differences in recovered views: Even when missing views are reconstructed, they may still contain subtle differences from the original data due to noise or reconstruction artifacts, which can lead to misclassifications.
The paper aims to address these limitations by developing a robust IMVC method that can handle fully incomplete information, where any view might be missing for any sample.
Methods
The proposed framework, FCITAFR (fuzzy clustering combined with information theory arithmetic based on feature reconstruction), tackles IMVC through a two-module approach:
- Feature Reconstruction Prediction Module:
- Within-View Reconstruction: Autoencoders are used to extract latent representations (Z) from the available views.
- Feature Reconstruction Prediction: This is the core novelty. Instead of directly predicting missing features, FCITAFR first reconstructs the features of existing views. It does this by:
- Calculating the average of features across all samples for each view.
- Using an entropy weighting method to analyze these latent features, assigning weights to different features based on their importance.
- Multiplying these weights with the initial latent representations to generate new, enhanced latent representations (L). This step aims to capture more robust and discriminative features from the available data.
- Missing View Prediction: Generators (G) then use these reconstructed features (L) to predict the missing views.
- Fuzzy Clustering Contrast Learning Module:
- Variational Fuzzy C-Means (FCM) Clustering: This is combined with information theory to learn the relationships and mutual information among the different views.
- Similarity Matrix Calculation: Similarity matrices (S) are computed for each view based on the reconstructed features and membership matrices (U) derived from FCM.
- Information Theory Integration: The similarity matrices are further enhanced by integrating information theory to capture cross-view joint probability distributions, leading to a fusion similarity matrix. This module aims to learn a consistent representation across views.
The overall objective function is a weighted sum of three loss components: fuzzy clustering with information theory loss ($L_{fcit}$), within-view reconstruction loss ($L_{rec}$), and feature reconstruction prediction loss ($L_{pre}$).
Datasets
The experiments were conducted on two multiphase liver tumor CECT datasets:
- Zhongda Dataset: An internal dataset from Zhongda Hospital, Southeast University, comprising 82 patients and 328 CT scans.
- Zheyi Dataset: An external dataset from The First Affiliated Hospital of Zhejiang University, with 475 patients and 1900 CT scans.
Both datasets were preprocessed to retain instances with microvascular invasion (MVI) annotations, a key indicator for postoperative liver diagnosis. The non-contrast (NC) and arterial phase (AP) were used as the two views for the experiments, with varying missing rates (10%, 30%, 50%, 70%, and 100%) simulated to evaluate the method's robustness.
Key Insights and Ideas
- Reconstruction Before Prediction is Crucial: The paper strongly advocates for reconstructing existing features before attempting to predict missing ones. This approach helps to consolidate and enhance the most reliable information available, rather than relying solely on potentially noisy or incomplete direct predictions.
- Information Theory for Cross-View Consistency: Integrating information theory with fuzzy clustering is a powerful way to quantify and leverage the mutual information between different views, leading to a more cohesive and accurate clustering.
- Robustness to High Missing Rates: The proposed FCITAFR method demonstrates remarkable resilience to high percentages of missing data, outperforming other methods when a significant portion of the information is absent. This is particularly relevant for real-world clinical scenarios where complete datasets are rare.
- Feature Reconstruction Loss is Key: Ablation studies highlight that the feature reconstruction loss ($L_{pre}$) plays a significant role, and its absence leads to a substantial drop in performance, underscoring the importance of the proposed reconstruction mechanism.
Key Results
FCITAFR consistently outperformed state-of-the-art IMVC methods across both datasets and various missing rates. Key quantitative results include:
- Accuracy (ACC): Achieved up to 77.5% on Zhongda and 79.6% on Zheyi at 10% missing rate. Even at 70% missing rate, it reached 72.4% (Zhongda) and 63.6% (Zheyi).
- Normalized Mutual Information (NMI): Reached 37.9% on Zhongda and 31.0% on Zheyi at 10% missing rate.
- Adjusted Rand Index (ARI): Achieved 29.5% on Zhongda and 34.6% on Zheyi at 10% missing rate.
- Robustness: At 100% missing rate (where only one view is available for each sample), FCITAFR achieved 58.6% ACC on Zhongda and 55.0% on Zheyi, significantly outperforming DCP.
- Efficiency: The method is noted as being remarkably efficient compared to some competitors, saving time on both datasets.
Limitations and Future Work
The authors acknowledge a few limitations and suggest future directions:
- Two-View Limitation: The current framework is designed for two views. Future work aims to extend FCITAFR to handle scenarios with more than two views, which are common in clinical practice.
- Unaligned Samples: The current method is primarily tested on aligned samples. Extending it to handle unaligned samples is another important future direction.
- Computational Cost: While efficient compared to some, the method's computational cost for very large datasets or more than two views might still be a consideration.
- Hyperparameter Tuning: The paper analyzes the impact of trade-off parameters ($\lambda_1, \lambda_2$) and suggests optimal values, but fine-tuning these might still be necessary for specific applications.
Summary updated