Multi-layer feature aggregation network with residual module and attention mechanism for jaw cyst image segmentation

Area of Science:

• Medical image analysis
• Computer-aided diagnosis
• Deep learning for healthcare

Background:

• Medical image segmentation is crucial for early disease detection and treatment planning.
• Existing methods struggle with blurred boundaries, low contrast, and varied lesion morphology.
• Challenges include enlarging receptive fields and extracting discriminative features.

Purpose of the Study:

• To propose MFA-Net, an innovative architecture for accurate medical image segmentation.
• To address limitations in feature extraction and receptive field size in current segmentation models.
• To improve the segmentation of challenging medical images, including lesions with indistinct characteristics.

Main Methods:

• Developed MFA-Net with specialized residual modules using 3x1/1x3 and 9x1/1x9 convolutional kernels.
• Incorporated a channel-spatial attention mechanism in the final encoder layer for enhanced feature representation.
• Utilized multi-layer feature aggregation by combining feature maps from adjacent encoder units for multi-scale feature synergy.

Main Results:

• MFA-Net achieved superior performance on jaw cyst datasets (F1: 0.9411, IoU: 0.9444) and the ISIC-2018 dataset (F1: 0.8678, IoU: 0.8388).
• Outperformed state-of-the-art methods across multiple evaluation metrics, including F1, IoU, Mcc, and Jaccard index.
• Ablation studies confirmed the significant contributions of the residual structure, attention mechanism, and feature aggregation block.

Conclusions:

• MFA-Net demonstrates significant improvements in medical image segmentation accuracy and robustness.
• The proposed architecture effectively addresses challenges related to feature extraction and receptive field limitations.
• MFA-Net shows strong potential for clinical applications in disease detection and treatment planning.