CD-Net: Cascaded 3D Dilated convolutional neural network for pneumonia lesion segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Accurate segmentation of COVID-19 lesions in Computer Tomography (CT) scans is vital for patient monitoring.
• Challenges in COVID-19 diagnosis include limited labeled data and redundant parameters in 3D CT scans.

Purpose of the Study:

• To develop an efficient model for segmenting COVID-19 lesions in CT scans.
• To address the limitations of data scarcity and computational complexity in existing methods.

Main Methods:

• A novel Cascaded 3D Dilated convolutional neural network (CD-Net) was developed to process CT volume data directly.
• A cascade architecture was designed to preserve global information and reduce memory consumption.
• A Multi-scale Parallel Dilated Convolution (MPDC) block was incorporated to aggregate features and reduce parameters.
• Transfer learning was employed to mitigate the shortage of labeled data.

Main Results:

• CD-Net demonstrated superior performance in segmenting COVID-19 lesions compared to existing methods.
• The model achieved better results with limited labeled data, indicating the effectiveness of transfer learning.
• CD-Net successfully reduced the negative-positive ratio in segmentation tasks.

Conclusions:

• The proposed CD-Net offers an effective solution for COVID-19 lesion segmentation in CT scans.
• CD-Net's architecture addresses challenges related to data scarcity and computational efficiency.
• This model shows significant potential for improving the screening and monitoring of COVID-19 cases.