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

Abstract

Nasopharyngeal carcinoma (NPC) is the most common malignant tumor of the nasopharynx. The delicate nature of the nasopharyngeal structures means that noninvasive magnetic resonance imaging (MRI) is the preferred diagnostic technique for NPC. However, NPC is a typically infiltrative tumor, usually with a small volume, and thus, it remains challenging to discriminate it from tightly connected surrounding tissues. To address this issue, this study proposes a voxel-wise discriminate method for locating and segmenting NPC from normal tissues in MRI sequences. The located NPC is refined to obtain its accurate segmentation results by an original multiviewed collaborative dictionary classification (CODL) model. The proposed CODL reconstructs a latent intact space and equips it with discriminative power for the collective multiview analysis task. Experiments on synthetic data demonstrate that CODL is capable of finding a discriminative space for multiview orthogonal data. We then evaluated the method on real NPC. Experimental results show that CODL could accurately discriminate and localize NPCs of different volumes. This method achieved superior performances in segmenting NPC compared with benchmark methods. Robust segmentation results show that CODL can effectively assist clinicians in locating NPC.

Figure 1

Example MR slices with three sequences. From left to right: (a) ground truth, (b) T1-w, (c) T2-w, and (d) CET1-w.

Figure 2

The workflow of location and segmentation of NPC. It consists of two steps, rough location by FCN and pixel-wise fine classification by CODL.

Figure 3

A toy example to demonstrate the discrimination power of the CODL. The data is collected from three views on (a) X-Y plane, (b) Y-Z plane, and (c) X-Z plane. The reconstructed results in X-Y-Z space by CODL on (d) the intact noiseless data, (e) the noisy data with std = 0.5, and (f) its std = 1 noisy counterpart are also shown. Different classes are highlighted in different colors.

Figure 4

NPC segmentation results on three typical examples. (a) Rough location results with bounding boxes identified by FCN, highlighted in red dots. The extended areas used for fine classification were indicated by solid red lines. (b) Fine segmentation results with fusing T1-w, T2-w, and CET1-w MR sequences. The last three columns are the tumor regions located by MISL, MvDA-VC, and CODL, respectively. (c) Results of our method on the whole slice in case of a combination of three modalities.

Figure 5

Typical segmentation results of three instances using CODL. (a) Ground truth. From second to last column: the tumor regions identified by CODL on modality T1-w (b), T2-w (c), CET1-w (d), both T1-w and T2-w (e), and T1-w, T2-w, and CET1-w (f), respectively.

Figure 6

Quantitative results of MISL, MvDA-VC, and CODL on multiple sequences of (a) T1-w and T2-w and (b) T1-w, T2-w, and CET1-w.

Figure 7

NPC segmentation results by fusing T1-w and T2-w modalities on the whole slices. (a) Ground truth. From second to last column: the tumor regions located by (b) Zhao et al. [12], (c) Li et al. [13], and (d) our approach, respectively.

Figure 8

NPC segmentation results by fusing T1-w, T2-w, and CET1-w modalities on the whole slices. (a) Ground truth. From second to last column: the tumor regions located by (b) Zhao et al. [12], (c) Li et al. [13], and (d) our approach, respectively.

Figure 9

Performance of our model on NPC dataset with different parameter settings by fusing T1-w and T2-w modalities: (a) hyperparameter λ₁ and (b) hyperparameter λ₂.

Figure 10

Performance of our model on the NPC dataset with different parameter settings by fusing T1-w, T2-w, and CET1-w modalities: (a) hyperparameter λ₁ and (b) hyperparameter λ₂.

Algorithm 1

The algorithm for solving CODL.