Research on a Multiscale U-Net Lung Nodule Segmentation Model Based on Edge Perception and 3D Attention Mechanism Improvement
Lung nodule semantic segmentation using deep learning has achieved good results. However, problems such as information loss on lesion edges, boundary segmentation blurring, lung nodule misdection, and low segmentation accuracy remain in lung CT (Computed Tomography) detection using deep learning due...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10747358/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Lung nodule semantic segmentation using deep learning has achieved good results. However, problems such as information loss on lesion edges, boundary segmentation blurring, lung nodule misdection, and low segmentation accuracy remain in lung CT (Computed Tomography) detection using deep learning due to the high degree of heterogeneity and a wide variety of nodule sizes, shapes, and locations, as well as the characteristics of convolutional localized feature extraction and the limitations of the continuous downsampling receptive field. So, a new model called EMC-UNet (Edge-aware _ Multiscale feature extraction residual _ 3D CA-Net attention module _ 3D U-Net), which integrates edge-awareness, 3D attention (3D CA-Net, Three-dimensional coordinate attention mechanism network), and multiscale techniques for segmenting lung nodules, is introduced. The model first uses an edge-aware module to accurately locate lesion edges, extract key edge features in the image, and increase the perception of lesion edge features by the model. Then, a 3D attention mechanism is added to focus the network on important lesion image features, emphasizing that the lesion features can improve segmentation performance. In conclusion, the 3D multiscale feature extraction module enhances the network’s perceptual range by processing information at various scales simultaneously, capturing features to offer a more comprehensive object context. This approach achieves notable results, with a Dice coefficient of 87.95% and an IoU value of 78.5% on the publicly available LIDC-IDRI(The Lung Image Database Consortium and Image Database Resource Initiative) dataset, outperforming existing lung nodule segmentation models. |
|---|---|
| ISSN: | 2169-3536 |