CA²PNet: a context-aware multi-scale architecture with adaptive attention and progressive dilated convolutions for...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Accurate medical image segmentation is crucial for computer-aided diagnosis but is hindered by data complexity, patient variability, and irregular patterns.
• Existing segmentation models often struggle with preserving fine details and global context simultaneously.

Purpose of the Study:

• To introduce a novel Context Aware Adaptive Progressive Network (CA²PNet) for enhanced medical image segmentation.
• To address limitations in current segmentation architectures, particularly resolution loss and boundary adherence.

Main Methods:

• The proposed CA²PNet integrates spatial attention modules (SAM), Global Max Pooling, enhanced Spatial Pyramid Pooling, and progressive dilated convolutions.
• This architecture refines local feature extraction while preserving global context, inspired by DeepLabV3+ and FusionNet.

Main Results:

• CA²PNet achieved a mean intersection over union of 85.15% on the Kvasir-SEG dataset and 82.78% on the BUSI dataset.
• The model outperformed existing state-of-the-art methods, with statistical tests confirming its robustness.

Conclusions:

• The CA²PNet effectively overcomes resolution loss in segmentation by embedding multi-scale features and employing decoupled decoding.
• This leads to superior boundary adherence and scale-invariant segmentation, improving diagnostic accuracy.