Automatically transferring supervised targets method for segmenting lung lesion regions with CT imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate segmentation of infected lung regions in CT images is crucial for diagnosis and treatment.
• Traditional methods often rely on extensive manual annotation, which is time-consuming and labor-intensive.
• Developing automated or semi-automated approaches can significantly improve efficiency and scalability.

Purpose of the Study:

• To present a novel semi-supervised dual-branch framework for autonomous identification and selection of optimal targets to segment infected lung regions in CT images.
• To enhance learning efficiency by utilizing both expert-annotated and automatically generated, coarsely annotated data.
• To improve the accuracy of deep learning models for medical image segmentation.

Main Methods:

• A semi-supervised dual-branch framework was designed, incorporating limited expert-annotated data and abundant coarsely annotated data (segmented using Hu values).
• The Lovasz scoring method was employed to dynamically switch and select optimal supervision targets within the weak branch during training.
• This approach allows the model to leverage noisy labels for initial localization and progressively refine targets using more accurate data.

Main Results:

• The proposed semi-supervised dual-branch network achieved a mean Dice Similarity Coefficient (DSC) of 83.56 ± 12.10% on internal benchmarks and 82.67 ± 8.04% on external benchmarks.
• Compared to U-Net without extra samples and the mean-teacher algorithm, the proposed method showed significant improvements in DSC values (up to 13.54% and 13.37% on internal and external benchmarks, respectively).
• The use of cost-effective pseudolabeled samples significantly boosted model performance.

Conclusions:

• Cost-effective pseudolabeled samples effectively assisted deep learning (DL) model training, outperforming traditional DL models trained solely on manual labels.
• The proposed method demonstrated superior performance compared to existing dual-branch structures.
• This approach offers a scalable and efficient solution for medical image segmentation tasks.