A Collaborative Dictionary Learning Model for Nasopharyngeal Carcinoma Segmentation on Multimodalities MR Sequences

Haiyan Wang, Guoqiang Han, Haojiang Li, Guihua Tao, Enhong Zhuo, Lizhi Liu, Hongmin Cai, Yangming Ou

TL;DR

This paper introduces a novel Collaborative Dictionary Learning (CODL) model for segmenting Nasopharyngeal Carcinoma (NPC) from multi-modal MRI sequences. By integrating information from different MRI views (T1-w, T2-w, and contrast-enhanced T1-w), CODL creates a discriminative latent space that significantly improves segmentation accuracy and robustness compared to existing methods.

What Problem Does This Paper Address?

Nasopharyngeal Carcinoma (NPC) is a challenging malignant tumor to detect and segment due to its infiltrative nature and small volume, making it difficult to distinguish from surrounding tissues in MRI scans. While MRI is the preferred diagnostic tool, accurately delineating NPC is crucial for treatment planning. Existing methods often struggle with the subtle boundaries and varying appearances of NPC across different MRI sequences, leading to difficulties in precise localization and segmentation. This paper aims to address this by developing a more robust and accurate segmentation method that leverages the complementary information present in multiple MRI modalities.

Methods

The proposed approach, Collaborative Dictionary Learning (CODL), is a two-stage process:

  1. Rough Tumor Localization: A Fully Convolutional Network (FCN) is used to identify a bounding box around the suspicious tumor area. This initial step helps to focus subsequent analysis and mitigate issues arising from class imbalance (small tumor volume within a large image slice).
  2. Pixel-wise Fine Classification: The CODL model then takes the localized region and performs fine segmentation. CODL is a multiview learning framework that integrates information from different MRI sequences (T1-w, T2-w, and CET1-w). It learns a shared, discriminative latent space by:
  3. Collaborative Dictionary Learning: Each view has its own dictionary, and these dictionaries work together to reconstruct the data in a common latent space.
  4. Supervised Membership: The model incorporates class labels to ensure the learned latent space is discriminative, meaning it can effectively separate tumor from non-tumor voxels.
  5. Optimization: The model is solved using an alternating direction method, iteratively updating the dictionaries, latent space representation, and classification coefficients.

Datasets

The study used a dataset of 24 patients with nonmetastatic NPC from the Sun Yat-sen University Cancer Center (SYSUCC). MRI scans were acquired using a 3.0-T scanner with T1-w, T2-w, and CET1-w sequences. The paper also mentions experiments on a synthetic dataset to test the model's discrimination power and robustness to noise.

Key Insights and Ideas

  • Multimodality Fusion is Key: Combining information from T1-w, T2-w, and CET1-w MRI sequences significantly improves segmentation accuracy, as each sequence offers different structural and contrast information.
  • Latent Space Discriminability: The core innovation of CODL is its ability to learn a latent space that is not only "intact" (capturing shared information across views) but also "discriminative" (explicitly designed to separate classes). This is achieved by incorporating class labels into the dictionary learning process.
  • Robustness to Noise and Missing Modalities: CODL demonstrates remarkable robustness to noise in synthetic data and maintains high performance even when one modality (CET1-w) is unavailable, suggesting its adaptability to real-world clinical scenarios where not all scans might be feasible.
  • Two-Stage Approach for Imbalance: The initial FCN-based localization step effectively addresses the challenge of small tumor volumes within large images, preventing false positives and improving the efficiency of the fine segmentation.

Key Results

  • Superior Segmentation Accuracy: CODL consistently outperformed benchmark methods (SVM, MISL, MvDA-VC) across various metrics (SENS, DICE, AUC, IoU, MPA, HD) on both synthetic and real NPC datasets.
  • High Performance with Three Modalities: When fusing T1-w, T2-w, and CET1-w, CODL achieved an AUC of 0.910 ± 0.048 and a DICE score of 0.823 ± 0.096 on the cropped NPC dataset.
  • Robustness to Noise: On synthetic data, CODL maintained the highest accuracy (74.9%) even with significant Gaussian white noise (std=1), outperforming other methods that showed a more pronounced performance drop.
  • Effective even with Two Modalities: When only T1-w and T2-w were used, CODL still achieved superior results, indicating that even a subset of modalities can yield excellent segmentation.

Limitations and Future Work

  • Data Availability: The paper states that the NPC image data cannot be made available due to patient privacy concerns, which limits external validation.
  • Parameter Sensitivity: While the paper shows that CODL is relatively insensitive to its hyperparameters (λ₁ and λ₂) within suitable ranges, further exploration of optimal parameter tuning for diverse datasets could be beneficial.
  • Future Work: The authors suggest that their method could be extended to other medical image segmentation tasks and potentially improved by incorporating more advanced deep learning architectures or regularization techniques. The current work focuses on voxel-wise classification, and future research could explore more sophisticated region-based segmentation strategies.

Summary updated

BibTeX
@inproceedings{wang2020collaborative,
  author = {Haiyan Wang and Guoqiang Han and Haojiang Li and Guihua Tao and Enhong Zhuo and Lizhi Liu and Hongmin Cai and Yangming Ou},
  title = {{A Collaborative Dictionary Learning Model for Nasopharyngeal Carcinoma Segmentation on Multimodalities MR Sequences}},
  booktitle = {Comput. Math. Methods Medicine},
  year = {2020},
  url = {https://openreview.net/forum?id=yr29LZNad7M},
}
Figure