A convolutional neural network combined with positional and textural attention for the fully automatic delineation of primary nasopharyngeal carcinoma on non-contrast-enhanced MRI

Abstract

Background: Convolutional neural networks (CNNs) have the potential to automatically delineate primary nasopharyngeal carcinoma (NPC) on magnetic resonance imaging (MRI), but currently, the literature lacks a module to introduce valuable pre-computed features into a CNN. In addition, most CNNs for primary NPC delineation have focused on contrast-enhanced MRI. To enable the use of CNNs in clinical applications where it would be desirable to avoid contrast agents, such as cancer screening or intra-treatment monitoring, we aim to develop a CNN algorithm with a positional-textural fully-connected attention (FCA) module that can automatically delineate primary NPCs on contrast-free MRI.Methods: This retrospective study was performed in 404 patients with NPC who had undergone staging MRI. A proposed CNN algorithm incorporated with our positional-textural FCA module (A proposed ) was trained on manually delineated tumours (M 1st ) to automatically delineate primary NPCs on non-contrastenhanced T2-weighted fat-suppressed (NE-T2W-FS) images. The performance of A proposed , three wellestablished CNNs, Unet (A unet ), Attention-Unet (A att ) and Dense-Unet (A dense ), and a second manual delineation repeated to evaluate human variability (M 2nd ) were measured by comparing to the reference standard M 1st to obtain the Dice similarity coefficient (DSC) and average surface distance (ASD). The Wilcoxon rank test was used to compare the performance of A proposed against A unet , A att , A dense and M 2nd .Results: A proposed showed a median DSC of 0.79 (0.10) and ASD of 0.66 (0.84) mm. It performed better than the well-established networks A unet [DSC =0.75 (0.12) and ASD =1.22 (1.73) mm], A att [DSC =0.75 (0.10) and ASD =0.96 (1.16) mm] and A dense [DSC =0.71 (0.14) and ASD =1.67 (1.92) mm] (all P<0.01), but slightly worse when compared to M 2nd [DSC =0.81 (0.07) and ASD =0.56 (0.80) mm] (P<0.001). Conclusions: The proposed CNN algorithm has potential to accurately delineate primary NPCs on noncontrast-enhanced MRI.

Radiomics for Discrimination between Early-Stage Nasopharyngeal Carcinoma and Benign Hyperplasia with Stable Feature Selection on MRI

“…Third, radiomics models have the weakness of requiring the segmentation of lesions, which could limit their value in aiding workflows. Regardless, we previously reported success in using a CNN to segment the primary tumor of NPC in T2w-fs [ 33 , 34 ], which can be expanded to segment any mucosal thickening (including BH) observed in T2w-fs. Combined with this CNN, a fully automatic radiomics method that is useful for clinical workflows can be developed from based on work reported in this study.…”

BSMM-Net: Multi-modal neural network based on bilateral symmetry for nasopharyngeal carcinoma segmentation

“…SI-Net, a variant of Unet, showed an improved performance by utilizing the information of the adjacent image and the high-risk primary tumor contour of the adjacent image ( Xue et al, 2020 ). To automatically delineate NPC lesions in MRI images, Wong et al (2021) used the texture and position information to weigh the channels of skipping features in an Unet-like network. Bai et al (2021) achieved the rough segmentation of NPC lesions in CT images using ResNeXt-50 Unet, which is constructed by replacing the encoder of Unet with ResNeXt-50.…”

CA‐Unet++: An improved structure for medical CT scanning based on the Unet++ Architecture

“…The reason is that the medical image data set is small and the images are relatively simple, which makes the five-layer structure designed by UNet just right. 32 Since the network structure of UNet is so outstanding in the field of medical image segmentation, the next thing we need to do is to improve the algorithm based on UNet. In the following years, many scholars have designed many excellent models, such as R2U-Net, 33 Attention-Unet.…”

“…At the same time, the five‐layer network structure designed by UNet also has great advantages for medical image segmentation. The reason is that the medical image data set is small and the images are relatively simple, which makes the five‐layer structure designed by UNet just right 32 …”