A weakly supervised learning-based segmentation network for dental diseases.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Supervised deep learning for medical image segmentation requires extensive labeled data.
• Existing methods suffer from private dataset bias, limiting model performance and generalization.
• Accurate segmentation is crucial for computer-aided medical diagnosis.

Purpose of the Study:

• To develop an end-to-end weakly supervised semantic segmentation network.
• To improve the robustness and generalization of medical image segmentation models.
• To enhance the accuracy of dental disease identification.

Main Methods:

• Proposed an attention compensation mechanism (ACM) using class activation maps (CAM) for complementary learning.
• Integrated conditional random fields (CRF) to refine foreground and background segmentation.
• Utilized high-confidence regions as pseudo-labels for training with a joint loss function.

Main Results:

• Achieved a Mean Intersection over Union (MIoU) of 62.84% in dental disease segmentation.
• Demonstrated an 11.18% improvement over previous networks.
• Verified enhanced robustness to dataset bias through an improved CAM localization mechanism.

Conclusions:

• The proposed weakly supervised network effectively segments dental diseases.
• The approach significantly improves accuracy and robustness compared to existing methods.
• This technique offers a promising solution for overcoming data limitations in medical image analysis.