Semantic segmentation of COVID-19 lesions with a multiscale dilated convolutional network

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate segmentation of COVID-19 lesions in CT scans is vital for diagnosis and monitoring.
• Challenges include scale variations and tissue similarity, hindering precise segmentation.

Purpose of the Study:

• To develop an automated method for segmenting COVID-19 lesions of varying scales and boundaries in CT images.
• To address challenges posed by scale differences and low contrast between lesions and normal tissues.

Main Methods:

• A novel multiscale dilated convolutional network (MSDC-Net) was proposed.
• The network incorporates a multiscale feature capture block (MSFCB) and a multilevel feature aggregate (MLFA) module.
• Utilized the publicly available COVID-19 CT Segmentation dataset for experiments.

Main Results:

• MSDC-Net outperformed existing methods in segmenting lesion boundaries and lesions of all sizes (large, medium, small).
• Achieved Dice similarity coefficient of 82.4%, sensitivity of 81.1%, and mean intersection-over-union (mIoU) of 78.2%.
• Demonstrated average improvements of 10.6% in Dice and 11.8% in mIoU compared to other methods.

Conclusions:

• The proposed MSDC-Net offers superior accuracy in segmenting diverse COVID-19 lesions and their boundaries on CT scans.
• This advancement facilitates improved clinical analysis and supports the development of automated COVID-19 diagnosis systems.