IAS-NET

IAS-NET performs unsupervised cross-domain adaptation for neonatal brain MRI segmentation by integrating a Generative Adversarial Network (GAN) with a U-net segmentation network to synthesize intraclass-aligned images and generate pseudo-labels for target-domain segmentation.

Key Features:

• Intraclassly Adaptive GAN (IA-NET): IA-NET implements a GAN with one generator (two encoders and a shared decoder) and four discriminators, where the dual encoders extract original image features and mean and variance from both source and target domains and a local adaptive instance normalization algorithm aligns intraclass features at the feature-map level.
• Segmentation Network (S-NET): S-NET uses a U-net architecture to provide semantic constraints via segmentation loss during IA-NET training and to generate pseudo-labels for target-domain images used to compute intraclass features.
• Unsupervised Domain Adaptation (UDA): IAS-NET performs UDA to bridge gaps between labeled source and unlabeled target domains and addresses regional misalignment and incorrect intensity information by enforcing intraclass similarity during image synthesis.
• Cross-domain Transfer: The method facilitates transfer from both image and feature perspectives to adapt image appearance and feature distributions across domains.
• Intraclass Feature Alignment: Intraclass alignment ensures synthesized images maintain class-specific characteristics within each tissue class to improve segmentation accuracy.

Scientific Applications:

• Neonatal brain MRI segmentation: IAS-NET targets segmentation of neonatal brain MR images by improving cross-domain generalization between different acquisition domains.
• NeoBrainS12 evaluation: On the NeoBrainS12 dataset, IAS-NET improved the average Dice score by 6% through adaptive training with unlabeled dHCP images.
• dHCP evaluation: On the dHCP dataset, IAS-NET achieved a 4% improvement in the average Dice score.

Methodology:

Generative Adversarial Network with one generator (two encoders and a shared decoder) and four discriminators; dual encoders extract original features and mean/variance for source and target; local adaptive instance normalization for intraclass feature alignment; U-net segmentation network providing segmentation loss and pseudo-labels; cross-domain transfer at both image and feature levels for unsupervised domain adaptation.