Depth-aware Network for Polyp Segmentation

Objectives Polyp segmentation is crucial for preventing colorectal cancer and diagnosing intestinal diseases. However, existing polyp segmentation algorithms often fail to consider the three-dimensional structure of polyps, limiting their performance. This study proposes a Depth-Aware Network (DANet) for polyp segmentation, integrating depth information to enhance segmentation accuracy, particularly in identifying polyp boundaries and shapes. The method is designed to improve the model’s capability in capturing the spatial structure of polyps, enhancing overall segmentation accuracy.Methods The proposed DANet consists of three main components: a spatial branch, a depth branch, and a feature fusion module. Specifically, the spatial branch utilizes a Pyramid Vision Transformer (PVT)-based encoder, which captures multi-scale feature representations from endoscopic images. This encoder processes the input image through multiple stages, progressively reducing the spatial resolution while extracting detailed and hierarchical features. This approach enables the network to capture local and global spatial information, which is essential for accurately segmenting polyps of varying sizes and shapes. The depth branch is responsible for generating depth maps that provide critical insights into the three-dimensional structure of polyps. These depth maps are obtained using the Depth Anything Model, a robust pre-trained model capable of producing high-quality depth estimates from endoscopic images. The generated depth maps provide valuable information regarding polyp protrusions and shapes, which are often challenging to discern using conventional 2D imaging techniques. A depth attention module is incorporated to enhance features specifically related to polyp protrusions. This module dynamically refines focus on depth-based features, enabling the network to emphasize regions with significant structural variations and improving segmentation accuracy. The feature fusion module integrates spatial and depth features through an adaptive fusion process. This fusion occurs across multiple scales, ensuring that depth-related variations are effectively incorporated into the spatial feature maps at different resolutions. The model successfully recognizes polyp boundaries and shapes by combining spatial and depth features. The adaptive fusion mechanism allows the network to selectively prioritize the most informative features from both domains, leading to more precise and robust segmentation results.Results and Discussions Experiments were conducted on five public polyp segmentation datasets, and DANet’s performance was compared to several state-of-the-art methods. The results indicated that DANet substantially improved segmentation accuracy across multiple evaluation metrics, including mean Dice (mDice) and mean Intersection over Union (mIoU). For example, on the Kvasir–SEG dataset, DANet attained a mDice score of 0.911 and a mIoU of 0.855, outperforming other methods such as PraNet, SANet, and MSNet. Similarly, on the CVC–ClinicDB dataset, DANet achieved a mDice of 0.934 and a mIoU of 0.884, demonstrating its superior performance. The incorporation of depth information enhanced the model’s ability to capture the three-dimensional structure of polyps, which is crucial for accurately identifying polyp boundaries and shapes. In addition, DANet exhibited strong generalization capabilities when evaluated on datasets such as ETIS and EndoScene, yielding competitive results across all metrics. Ablation studies were conducted to assess the contributions of each module. The results showed that the entire model, which includes both the depth branch and the feature fusion module, achieved the best performance, confirming the effectiveness of integrating depth information. Regarding computational efficiency, DANet maintains a balance between accuracy and processing speed. The network was trained using the PyTorch framework on an NVIDIA TITAN RTX 3090 GPU, completing the training process in 100 epochs. The results indicate that DANet is well-suited for real-time medical applications due to its efficient architecture.Conclusions The proposed Depth–Aware Network (DANet) significantly improves polyp segmentation accuracy by incorporating depth information. The integration of spatial and depth features enables the model to capture the three-dimensional characteristics of polyps more accurately, improving the recognition of polyp boundaries and shapes across different datasets. The experimental results validate DANet’s generalizability, while ablation studies confirm the importance of the depth and feature fusion modules. This approach has potential applications beyond polyp segmentation, extending to other medical image segmentation tasks that require accurate three-dimensional structural recognition.