BE-FNet: 3D Bounding Box Estimation Feature Pyramid Network for Accurate and Efficient Maxillary Sinus Segmentation

Abstract EN

Maxillary sinus segmentation plays an important role in the choice of therapeutic strategies for nasal disease and treatment monitoring.

Difficulties in traditional approaches deal with extremely heterogeneous intensity caused by lesions, abnormal anatomy structures, and blurring boundaries of cavity.

2D and 3D deep convolutional neural networks have grown popular in medical image segmentation due to utilization of large labeled datasets to learn discriminative features.

However, for 3D segmentation in medical images, 2D networks are not competent in extracting more significant spacial features, and 3D ones suffer from unbearable burden of computation, which results in great challenges to maxillary sinus segmentation.

In this paper, we propose a deep neural network with an end-to-end manner to generalize a fully automatic 3D segmentation.

At first, our proposed model serves a symmetrical encoder-decoder architecture for multitask of bounding box estimation and in-region 3D segmentation, which cannot reduce excessive computation requirements but eliminate false positives remarkably, promoting 3D segmentation applied in 3D convolutional neural networks.

In addition, an overestimation strategy is presented to avoid overfitting phenomena in conventional multitask networks.

Meanwhile, we introduce residual dense blocks to increase the depth of the proposed network and attention excitation mechanism to improve the performance of bounding box estimation, both of which bring little influence to computation cost.

Especially, the structure of multilevel feature fusion in the pyramid network strengthens the ability of identification to global and local discriminative features in foreground and background achieving more advanced segmentation results.

At last, to address problems of blurring boundary and class imbalance in medical images, a hybrid loss function is designed for multiple tasks.

To illustrate the strength of our proposed model, we evaluated it against the state-of-the-art methods.

Our model performed better significantly with an average Dice 0.947±0.031, VOE 10.23±5.29, and ASD 2.86±2.11, respectively, which denotes a promising technique with strong robust in practice.