Lesion segmentation in lung CT scans using unsupervised adversarial learning.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate lesion segmentation in medical images is vital for diagnosis and treatment but is often hindered by the scarcity of annotated data.
• Existing supervised and semi-supervised methods struggle with limited annotations, necessitating alternative approaches.

Purpose of the Study:

• To develop an unsupervised learning technique for automatic segmentation of coronavirus disease 2019 (COVID-19) lesions in 2D axial CT lung slices.
• To address the challenge of limited annotated data in medical imaging by utilizing unsupervised learning.

Main Methods:

• Proposed an unsupervised image translation technique to generate synthetic healthy lung images from infected ones, eliminating the need for lesion annotations.
• Incorporated attention masks to enhance the precision and quality of the lesion segmentation process.

Main Results:

• The unsupervised approach demonstrated capability in segmenting COVID-19 lesions, outperforming existing unsupervised lesion detection methods.
• Achieved an average Dice Score of 0.695, Sensitivity of 0.694, Specificity of 0.961, Structure Measure of 0.791, Enhanced-Alignment Measure of 0.875, and Mean Absolute Error of 0.082.

Conclusions:

• The proposed unsupervised learning method shows promise for automatic COVID-19 lesion segmentation in CT scans.
• This technique offers a valuable, annotation-free tool for medical image analysis and could advance future diagnostic developments.