U-SEANNet: A Simple, Efficient and Applied U-Shaped Network for Diagnosing Nasal Diseases from Nasal Endoscopic Images
share
Utilizing deep learning (DL) models to improve the early diagnosis of nasal diseases from nasal endoscopic images holds paramount importance. However, the lack of available datasets stymies advancements in this field. Furthermore, existing models fail to strike a good trade-off between model diagnosis performance, model complexity and parameter size, rendering them unsuitable for practical application. To bridge these gaps, we created the first large-scale nasal endoscopy dataset, named 7-NasEID, comprising 11,352 images that span six nasal diseases and normal samples. Building on this, we proposed U-SEANNet, an innovative architecture, underpinned by depth-wise separable convolutions. Additionally, to augment its discernment capabilities for subtle variations in input images, we further proposed the Global-Local Channel Feature Fusion Module, enabling the U-SEANNet to focus salient channel features from both global and local contexts. Notably, U-SEANNet's parameter size and GFLOPs are only 0.78M and 0.21, respectively. Employing the 7-NasalEID, we conducted the five-fold cross-validation on U-SEANNet, juxtaposing its performance against seventeen renowned architectures. The experimental results suggest U-SEANNet as the state-of-the-art (SOTA) model, achieves an accuracy of 93.58 of 90.17 prodigious potential for diagnosing nasal diseases in practical use, providing the development of efficacy nasal diseases diagnosis tools with a new insight.
READ FULL TEXT
