SAAMU-Net: A Selective Attention Aggregation and Memory-Augmented U-Net Model for Microscopic Medical Image Segmentation

This is the authors' abstract. We don't add key points for this paper.

Microscopic medical image segmentation is vital for detailed cellular analysis in computational pathology, yet it remains challenging due to overlapping structures and high intra-class variability. To address this, we propose SAAMUNet, a U-Net-based model with an InceptionV4 backbone that enhances segmentation accuracy through Selective Attention Aggregation (SAA) and Memory-Augmented Attention (MAA) modules. SAA refines spatial details by merging global and local cues, while MAA improves segmentation consistency by recalling past feature activations, ensuring enriched skip connections for robust feature continuity. We have applied Dual-Tree Complex Wavelet Transform (DTCWT) on the input image to enhance multi-scale feature representation. The encoder extracts hierarchical features with progressively increasing filters (64 to 1024), while the decoder restores spatial details through upsampling, skip connections, and attentionenhanced refinement. Multi-scale outputs and deep supervision further boost accuracy while maintaining computational efficiency. Extensive evaluations on TNBC, CPM17, and Cervical Nucleus Segmentation (CNS) datasets demonstrate the ability of SAAMU-Net to achieve high segmentation precision, highlighting its potential to advance computational pathology.