Attention Mechanism Trained with Small Datasets for Biomedical Image Segmentation

The understanding of long-range pixel–pixel dependencies plays a vital role in image segmentation. use CNN plus an attention mechanism still has room for improvement, since existing transformer-based architectures require many thousands annotated training samples to model spatial dependencies. This paper presents smooth branch (SAB), novel architecture that simplifies the biomedical segmentation small datasets. SAB is essentially modified operation implements subnetwork via reshaped feature maps instead directly calculating softmax value over score each input. fuses multilayer attentive learn visual multilevel features. We also introduce position blurring and inner cropping specifically small-scale datasets prevent overfitting. Furthermore, we redesign skip pathway reduction semantic gap between every captured contracting expansive path. evaluate U-Net with (SAB-Net) by comparing it original widely used models across multiple tasks related Brain MRI, Heart Liver CT, Spleen Colonoscopy Our set was made random 100 images set, our goal adopt mechanisms labeled data. An ablation study conducted on brain MRI test demonstrated proposed method achieved improvement Integrating methods helped resulting consistently achieve outstanding performance above five tasks. In particular, improved its 13.76% dataset. several address need modeling experimental results illustrated could improve medical accuracy various degrees. Moreover, SAB-Net, which integrated all methods,