The aim of the study is to evaluate the performance of U-Net in tumor segmentation on diffusion MR imaging for patients with cervical cancer. Diffusion weighted imaging of b0, b1000 and ADC maps were used for training. The ADC histogram parameters of predicted region of tumor were assessed for accuracy and reproducibility. The results show the triple-channel training algorithm exhibited the best performance in both training and testing datasets. The predicted voxels of tumor can be used to generate the volumetric ADC data for Radiomics study.
Among the 169 patient databank, 144 patients were used for the training phase, while another independent 25 patients were used for the testing. The sagittal DW imaging of b0, b1000 and ADC maps of each slice of each patients were used as the input for the subsequent training.
The end-to-end neural network was adopted for tumor segmentation based on the U-Net architecture using a fully convolutional network 2. The architecture combined a contracting (down-sampling) path to capture context and a symmetric expending (up-sampling) path that enables precise localization 3. A total of 17 convolutional layers were employed in the network. Image augmentation was performed with cropping, rotation, and shifting, which generated 61320 images for each data sets. Seven combinations from the 3 image sets (b0, b1000 and ADC) were used as the input sources in different channels for training.
To test the reproducibility of the training model, a repeat training procedure was performed using the identical parameters. After building the models, we examined the accuracy of the models using the testing dataset images. The accuracy of segmentation was evaluated using Dice similarity coefficient (DSC),. The histogram parameters of ADC values (mean, minimum, 10th, 25th, 50th, 75th, 90th percentiles, maximum, kurtosis, skewness and standard deviation) extracted by the predicted algorithm was correlated with those from manually labeled using Pearson correlation. The reproducibility of the training was assessed using the Intraclass Correlation Coefficient (ICC) between the ADC histogram parameters of ROIs extracted from the 1st and 2nd training algorithms.
Performance in Training phase. In the training stage, The use of triple channel input (ADC+b0+b1000) exhibited the fastest learning efficacy to reach the plateau with the highest accuracy of 0.946. The use of single channel of b0 had the lowest learning efficacy with the plateau accuracy of 0.891 (Fig1).
Performance of in validation phase. Fig.2 illustrates an example of fully automated tumor segmentation for a patient with cervical cancer in the testing dataset. The predicted ROI, which was generated using the triple-channel training algorithm (ADC + b0 + b1000), was segmented correctly at the corresponding tumor location as in the labeled image. Fig.3 plots the accuracy of segmentation for various combinations of source images that fed to training, the triple channel model exhibited the highest DSC (0.95±0.05 and 0.82±0.07 respectively for the training and testing datasets).
Tumor ADC Histogram parameters. All the ADC histogram parameters of the tumor were highly correlated between the labeled and predicted ROIs. (r = 0.25 – 0.9, p <0.001 for all). The reproducibility of the histogram parameters between the 1st and 2nd trainings were high for all the histogram parameters (ICC = 0.79 – 0.98).