Related Experiment Video
Updated: May 7, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Attention mechanisms in UNet variants for medical/non-medical image segmentation: A comprehensive and
Siddharth Gupta1, Yogendra Chhetri2, Amit Pandey3
1Stroke Diagnostics and Monitoring Division, AtheroPoint™, Roseville, CA, 95661, USA.
Background:
This review examines enhancements to the U-Net's ability to represent complex spatial structures for segmentation through attention mechanisms. The study surveys applications across both medical and non-medical domains, focusing on cSAM, cCAM, CBAM, and their associated serial (sfSCAM) and parallel (pfSCAM) fusion architectures. Beyond accuracy, the analysis also considers backbone design, diversity of evaluation metrics, pruning strategies, and potential sources of bias.
Methodology:
Following PRISMA guidelines, we systematically searched major academic databases using U-Net- and attention-related keywords. The final corpus was constructed through stepwise exclusions (E1-E3) from an initial set of 1807 retrieved records. Spatial, channel, and hybrid attention modules were taxonomized; their placement within encoder, decoder, skip, and bottleneck blocks was identified; and application trends across domains such as liver, brain, skin, and retinal imaging were summarized.
Findings:
Attention-augmented U-Net models consistently improved the segmentation of subtle, scale-dependent structures across modalities, including CT, MRI, ultrasound, dermoscopy, fundus, and satellite imaging. Spatial attention in skip connections and decoders reduced noise while preserving boundaries, whereas encoder-level channel attention improved semantic weighting; CBAM combined both effects. Parallel fusion (pfSCAM) balanced accuracy and efficiency, while serial fusion (sfSCAM) enabled stable feature refinement. Attention placement strongly influenced global-local fusion and noise suppression. Pruning reduced computational cost with minimal performance loss, although evaluation remained focused on Dice and cross-entropy, with limited use of IoU, Hausdorff distance, and F1-score.
Conclusion:
Spatial and channel attention, along with their serial and parallel fusion variants, function as structural design principles within the U-Net. Effective practice integrates careful module placement, pruning or quantization strategies, and explicit bias auditing.
