Dental Lesion Segmentation Using an Improved ICNet Network with Attention.

Area of Science:

• Computer Science
• Biomedical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of dental lesions is crucial for developing intelligent diagnostic systems.
• Distinguishing tooth lesions from normal tissue presents a significant segmentation challenge.

Purpose of the Study:

• To propose an improved Image Cascade Network (ICNet) for accurate and real-time segmentation of diverse tooth lesions.
• To enhance feature relevance and suppress irrelevant information for improved segmentation accuracy.

Main Methods:

• Implemented a modified ICNet incorporating the Convolutional Block Attention Module (CBAM).
• Replaced large convolutions with layered dilated convolutions within the spatial attention module.
• Utilized asymmetric convolutions to optimize computational efficiency.

Main Results:

• The proposed method demonstrated superior segmentation performance compared to FCN, U-Net, and SegNet.
• Achieved higher image processing frequencies, meeting real-time requirements for dental lesion segmentation.
• Successfully segmented various lesion types including calculus, gingivitis, and tartar.

Conclusions:

• The enhanced ICNet with CBAM offers a significant improvement in tooth lesion segmentation accuracy and speed.
• This method provides a robust solution for intelligent tooth lesion detection systems.
• The approach effectively addresses the challenge of segmenting similar-looking tooth lesions and healthy tissues.