Dual-Attention-Based Enhanced Unified Net for Precise GTV Segmentation of Nasopharyngeal Carcinoma in 3D MR Images

Accurate gross tumor volume (GTV) segmentation is essential for effective radiotherapy in nasopharyngeal carcinoma (NPC). However, challenges arise due to the nasopharyngeal region’s complex anatomy and the annotated data scarcity. Our study presents a dual-attention-based enhanced unified network (DAEU-Net) designed for precise NPC GTV segmentation utilizing 3D T1, T2, and T1C-weighted MR images. Our approach involves splitting large-scale MR data into multiple patches and then training every patch independently. This approach effectively captures localized and detailed information without downscaling the image resolution. The DAEU-Net integrates channel-attention and pixel-attention modules within the encoder section, eliminating background noise and reducing information loss by enhancing the network’s focus on detailed features. The decoder section incorporates bottleneck residual blocks to enhance the computing efficiency and robustness of the network. The proposed methodology surpasses the state-of-the-art models with a respective average symmetric surface distance (ASSD) of 0.920±0.386 mm, 0.987±0.421 mm, and 1.043±0.457 mm and a dice similarity coefficient (DSC) of 0.896, 0.871, and 0.851, respectively. Multi-viewed animated MR images in three orthogonal dimensions (axial, sagittal, and coronal) with predicted NPC tumors and real GTV masks were shown to assist in comprehending the tumor’s precise location. Our approach can significantly improves NPC tumor delineation, aids in automated tumor lesion segmentation, and reduces the annotation workload for oncologists.