Self-Attentive Spatial Adaptive Normalization for Cross-Modality Domain Adaptation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks struggle with domain generalization in medical imaging.
• Cross-modality medical data presents significant domain shift challenges.
• Limited availability of annotated medical imaging data hinders model development.

Purpose of the Study:

• To develop a novel image-to-image translation method for medical images.
• To reduce the annotation burden for radiologists.
• To bridge the domain gap in radiological images for improved segmentation and conversion.

Main Methods:

• Cross-modality synthesis using adversarial training.
• A learnable self-attentive spatial normalization in deep convolutional generator networks.
• Incorporation of auxiliary semantic information to preserve anatomical structures and handle geometric changes.
• Supervised and unsupervised (unpaired data) image-to-image translation setups.

Main Results:

• Superior cross-modality segmentation results for unpaired MRI and CT data (whole heart, brain tumor).
• Encouraging cross-modality conversion results for paired MRI and CT brain images.
• Demonstrated efficacy in handling geometric changes and preserving anatomical structures.

Conclusions:

• The proposed method effectively performs cross-modality medical image translation.
• The approach reduces the need for costly annotations and improves generalization across domains.
• Self-attentive spatial normalization and auxiliary semantic information are crucial for robust medical image translation.