Table 1 Comparison of methods.

From: Automatic segmentation of karst reservoir CT images and identification of karst spatial structure based on 3D U-Net

Feature/method

Standard 3D U-Net

Attention U-Net

Multi-scale pyramid networks

Proposed method

Multi-scale input

Single-scale

Single-scale

Serial or fused input

Parallel multi-scale input paths

Attention mechanism

None

Channel/spatial attention

Typically integrated into encoder

Spatial attention embedded in skip connections

Core design objective

General segmentation

Enhance response to key regions

Enhance multi-scale perception

Preserve karst topological connectivity

Boundary optimization strategy

Dice/cross-entropy

Same as left

Same as left

Dice + boundary-aware loss (gradient L1)

Geological adaptability

Limited

Limited

Moderate

High (targeting karst scale and connectivity)