DA-DSUnet: Dual Attention-based Dense SU-net for automatic head-and-neck tumor segmentation in MRI images

Abstract

Precise and accurate segmentation of the most common head-and-neck tumor, nasopharyngeal carcinoma (NPC), in magnetic resonance images (MRI) sheds light on treatment and regulatory decisions making, especially on radiotherapy planning. However, segmenting NPC manually is time-consuming and expensive. Thus, automatic segmentation of NPC is highly demanded. Whereas, problems such as large variations in the lesion size and shape of NPC, boundary ambiguity, as well as limited available annotated samples, conspire NPC segmentation towards a formidable task. Consequently, existing NPC segmentation methods cannot satisfy the high requirements for medical practice. Motivated by these challenges and prevalence of deep learning, in this paper, we propose a Dual Attention-based Dense SU-net (DA-DSUnet) framework for automatic NPC segmentation. It is an encoder-decoder network taking 2D NPC MRI slices as input and outputting the corresponding segmentation results. The main innovations of our method are fourfold. First, different from the traditional decoder in the baseline (U-net) which uses upconvolution for upsamling, we argue that the restoration from low resolution features to high resolution outputs should be capable of preserving the information related to boundary localization. Therefore, we use unpooling as the upsampling method in our model. Second, to combat the vanishing-gradient problem, we introduce dense block which can facilitate feature propagation to replace the traditional convolutional block. Third, we incorporate a dual attention mechanism in our network, which models the inter-dependencies in position and channel dimensions. Fourth, using only binary cross entropy (BCE) as loss function may bring about troubles such as miss-prediction. Hence, we propose to use a loss function named BCEDice to train the network. Quantitative and qualitative comparisons are carried out extensively on in–house dataset. The experimental results show that the proposed method achieves a DSC of 0.8050, a PM of 0.8026 and a CR of 0.7065, which respectively has a relative gain of 5.17%, 13.8% and 10.3% over U-net, indicating the effectiveness of our method.