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

Accurate medical image segmentation remains a challenging task in computer-aided diagnosis because of the intricacies and the variability in the biomedical data in terms of the anatomical complexity, inter-patient diversity, class imbalance, and irregular morphological patterns.

In the present work, a Context Aware Adaptive Progressive Network (CA²PNet) is proposed. The foundational architecture of CA²PNet is inspired from DeepLabV3+ and FusionNet and introduces four key modifications by the incorporation of (a) the spatial attention module (SAM) to emphasize discriminative spatial regions, (b) Global Max Pooling to strengthen contextual representation and suppress background noise, (c) an enhanced Spatial Pyramid Pooling for robust multi-scale feature extraction, and (d) progressive dilated convolutions to expand the receptive field while preserving fine structural details. The incorporation of these modules enables the simultaneous refinement of the extraction of local features as well as preserving the global context.

CA²PNet offered a mean intersection of union for 85.15 and 82.78% for the Kvasir-SEG and BUSI datasets, respectively, surpassing state-of-the-art models. The statistical tests also validate the robustness of the proposed model.

The proposed work demonstrates that the embedding of multi-scale features at each encoder stage along with decoupled decoding aids in overcoming the limitation of resolution loss in classical segmentation architectures; thus, resulting in superior boundary adherence and scale-invariant segmentation.