The study presents an automatic recognition method for parotid gland tumor. We used a convolution neural network to conduct the segmentation of parotid gland tumor and classifications of tumor types. We also designed eight combinations of various MRI contrasts to compare the results of recognition for parotid gland tumor. We compared results obtained using various combinations of MR images as the input of the convolutional neural network and found that diffusion-related parameters and contrast-enhanced T1 images played the primary role of the prediction accuracy.
This study was approved by the institutional review board and written informed consents were waived. We collected 133 MRI datasets from 40 PGT patients (benign PGT: 23, malignant PGT: 17). Each dataset consisted of five types of multi-slice MR images encompassing the PGT region, including T2, T1 with contrast-enhancement, diffusion-weighted echo-planar images (DW-EPI, b values: 0, 1000 sec/mm2) and the corresponding apparent-diffusion-coefficient (ADC) maps. The five types of images are referred to as T2, T1ce, b0, bk, and ADC respectively. The T2, T1ce and DW-EPI images underwent co-registration in SPM software package. Also, we manually outlined the tumor regions and labeled the pixels as 0: background, 1: benign, and 2: malignant PGTs. We generated 8 combinations of the 5 types of MRI images as the input multi-channel images for the following deep-learning procedures. They were listed in Table 1. We normalized the images with the maximum intensity of each channel and padded zeros to each image to construct a multi-channel image training image (256×256×4). We distributed the subjects into three groups to conduct 3-fold cross-validation of deep learning.
We used two-dimensional SegNet for pixel-wise semantic segmentation.3 The architecture of SegNet implementation is displayed in Figure 3. In this study, we adapted a model won the third prize in BraTS 2017,4 as the pre-trained model. A four-channel input layer was implemented, and the SegNet was pre-trained with four types of brain MR images (T2, FLAIR, T1, T1ce). Approximately 2×105 2D images extracted from 285 3D datasets were used to train the BraTs model. We transferred the weights of the pre-trained model and trained a new model with the PGT datasets. The pre-initialization of the SegNet allowed the training procedure of the new PGT model converged to a moderate accuracy. Finally, we calculated the values of sensitivity, precision, and F1-score of the neural network and compare 8 types of multi-channel images.
Figure 2 displays the example results of recognition for PGT and Table 1 shows the group statistics of the accuracy indexes of PGT segmentations and classifications. Notice that the Comb-6 (b0 + bk + ADC + T1ce) performs best in the segmentation task according to the F1 score, which can also be identified in Figure 2. However, the classification of Comb-6 is not as accurate as the rest combinations. Thus, we fused two models for the tumor classifications, i.e., Comb-6 as the segmentation model and Comb-3, 4, or 5 as the classification model. The F1-scores of tumor classifications of Comb-6+Comb-4 and Comb-6+Comb-5 were close to 0.87.
We then merged the multi-slice results (133 slices) into 40 patients by a voting mechanism. All the tumors identified from the slices acquired from the same subject were collected, and the numbers of malignant and benign tumors were calculated to “vote” the malignancy of the PGT of this subject. In the event of a tie, the PGT was identified as malignant. Table 2 lists the accuracy metrics of identifying malignant PGTs by three configurations of input datasets. They were Comb-2 requiring only T2 and T1ce, Comb-3 requiring only b0, b0 and ADC obtained by DW-EPI, and Comb-6+Comb-4 requiring T1ce and DW-EPI-related images. The accuracy obtained using Comb-6+Comb-4 datasets is the highest amount the three methods.
1. Christe A, Waldherr C, Hallett R, Zbaeren P, Thoeny H, MR Imaging of Parotid Tumors: Typical Lesion Characteristics in MR Imaging Improve Discrimination between Benign and Malignant Disease, American Journal of Neuroradiology August 2011, 32 (7) 1202-1207.
2. Liu YJ, Lee YH, Chang HC, Huang TY, Chiu HC, Wang CW, Chiou TW, Hsu K, Juan CJ, Huang GS, and Hsu HH, A Potential Risk of Overestimating Apparent Diffusion Coefficient in Parotid Glands, PLoS One. 2015; 10(4): e0124118.
3. Badrinarayanan V, Kendall A, Chipolla R, SegNet: A Deep Convolutional Encoder-Decoder Architecture for Scene Segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, Jan 2017.
4. Yang TL, Ou YN, Huang TY, Automatic segmentation of brain tumor from MR images using SegNet: selection of training data sets, third prize, MICCAI 2017: Multimodal Brain Tumor Segmentation Challenge 2017