Multi-planar 2D-U-Net Segmentation of 3D-CT Abdominal Organs augmented by Spatial Occurrence Maps

Researchers propose a lightweight 2D-U-Net framework for segmenting abdominal organs in 3D CT scans by combining multi-planar analysis with spatial occurrence maps. The two-stage approach achieves approximately 4% Dice improvement over baseline models and demonstrates practical viability for medical imaging applications.

This research addresses a fundamental challenge in medical imaging: accurate automated segmentation of abdominal organs in volumetric CT data. The proposed framework leverages 2D convolutional neural networks applied across multiple anatomical planes rather than attempting full 3D processing, a design choice that reduces computational overhead while maintaining segmentation quality. The integration of spatial occurrence maps—fuzzy 3D probability distributions indicating where organs typically appear—provides anatomical context that guides the segmentation process and prevents anatomically implausible predictions.

The method's two-stage architecture reflects practical constraints in medical imaging. The initial coarse-to-fine detection stage efficiently identifies the region of interest across the entire scan volume, reducing unnecessary computation in subsequent fine-grained segmentation. By augmenting the multi-planar predictions with spatial priors, the framework achieves a meaningful 4% improvement in Dice coefficient, a standard metric for segmentation accuracy. This improvement proves clinically relevant, as enhanced organ delineation directly impacts downstream diagnostic and surgical planning workflows.

The evaluation on 80 CT scans from diverse public sources suggests reasonable generalization potential, though broader validation across different imaging protocols and patient populations would strengthen clinical applicability. For medical AI developers, this work demonstrates that combining architectural efficiency with domain-specific anatomical knowledge yields tangible performance gains without requiring computationally expensive 3D convolutions. The lightweight nature of 2D-U-Net models makes this approach viable for deployment in resource-constrained clinical environments where rapid analysis is essential.

• →Multi-planar 2D-U-Net segmentation achieves 4% Dice improvement when augmented with spatial occurrence maps
• →Two-stage coarse-to-fine approach efficiently identifies organ regions before fine-grained segmentation
• →Lightweight 2D architecture reduces computational burden compared to full 3D convolution methods
• →Spatial anatomical priors prevent anatomically implausible segmentation predictions
• →Method evaluated on diverse public CT datasets with potential for clinical deployment