CAML-PSPNet: A Medical Image Segmentation Network Based on Coordinate Attention and a Mixed Loss Function

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning

Background:

• Segmentation tasks often struggle with fuzzy boundaries and small regions, impacting diagnostic accuracy.
• Existing methods like PSPNet, Deeplabv3, HrNet, and U-Net have limitations in handling these segmentation challenges.

Purpose of the Study:

• To propose a novel network, CAML-PSPNet, that improves segmentation accuracy for medical images, particularly for small targets and fuzzy boundaries.
• To enhance the precision of clinical diagnoses by addressing common segmentation inaccuracies.

Main Methods:

• Developed CAML-PSPNet, integrating a coordinate attention module for precise edge localization and a Mixed Loss Function (MLF) for small-target segmentation.
• Utilized MobilenetV2 as a lightweight backbone for efficient feature extraction, reducing model parameters and increasing computation speed.
• Validated the model on PrivateLT, Kvasir-SEG, and ISIC 2017 datasets, comparing its performance against established networks.

Main Results:

• CAML-PSPNet demonstrated significant improvements in visual effects and evaluation metrics compared to Deeplabv3, HrNet, U-Net, and PSPNet.
• Achieved average intersection rate increases of 2.84%-5.4% on lung cancer data, 3.1%-8.78% on Kvasir-SEG, and 0.71%-3.83% on ISIC 2017.
• Showcased superior boundary segmentation similarity to the gold standard and enhanced accuracy for small targets.

Conclusions:

• CAML-PSPNet effectively addresses limitations in medical image segmentation, particularly concerning fuzzy boundaries and small targets.
• The proposed network offers a more accurate and computationally efficient solution for segmentation tasks, aiding in improved clinical diagnosis.
• The integration of coordinate attention and a mixed loss function represents a significant advancement in segmentation network design.