SE–UNet: Spatial and Channel Refinement for Robust Multi–Anatomical Structure Segmentation in Abdominal Ultrasound

Automated segmentation of multiple anatomical structures in abdominal ultrasound could enhance clinical decision–making by providing consistent quantitative measurements for disease monitoring. However, ultrasound–specific artifacts including speckle noise, low contrast and weak boundaries challenge existing deep learning models. We propose SE–UNet, a novel architecture that integrates spatial attention gates and squeeze–and–excitation blocks to improve boundary localization and tissue discrimination. On a multi–anatomical dataset combining kidney and spleen ultrasound images (six tissue classes), SE–UNet achieved a mean Dice score of 0.7165, outperforming nine baseline methods. Clinical validation with an expert radiologist on 50 cases demonstrated 0.78 Dice agreement with manual annotations, with 88% of segmentations deemed clinically acceptable for measurement purposes. The model processes images in 0.028 seconds, supporting real–time clinical workflow integration. Our approach demonstrates how combined spatial and channel refinement addresses key ultrasound segmentation challenges with direct clinical applicability.