OBELISK-Net: Fewer layers to solve 3D multi-organ segmentation with sparse deformable convolutions

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep networks excel in image analysis but require manual tuning of convolutional filters.
• 3D fully-convolutional networks are memory-intensive and need large datasets, limiting medical applications.

Purpose of the Study:

• To propose a novel, efficient deep learning method for 3D medical image analysis.
• To reduce memory and parameter requirements in deep networks for segmentation tasks.

Main Methods:

• Introduced a novel method using trainable 3D convolution kernels that learn filter coefficients and spatial offsets.
• Developed the one binary extremely large and inflecting sparse kernel (OBELISK) filter.
• Applied the OBELISK filter within a deep network for 3D CT multi-organ segmentation.

Main Results:

• The OBELISK filter requires fewer trainable parameters and less memory compared to U-Net architectures.
• Achieved high-quality results on challenging 3D CT multi-organ segmentation tasks.
• Demonstrated that sparse deformable convolutions capture large spatial context efficiently.

Conclusions:

• Network depth is not always necessary for learning complex features.
• The proposed method offers potential for computer-assisted and image-guided interventions due to fast inference and improved segmentation.