Cross-domain attention-guided generative data augmentation for medical image analysis with limited data

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Traditional data augmentation in medical imaging is limited by insufficient data and annotations.
• Existing generative adversarial networks (GANs) struggle with cross-domain information and supervised guidance for segmentation.
• These limitations hinder diagnostic and segmentation performance in medical image analysis.

Purpose of the Study:

• To develop an attention-guided cross-domain tumor image generation model (CDA-GAN) for enhanced medical image data augmentation.
• To address the limitations of existing generative methods in utilizing cross-domain information and providing supervised signals for segmentation.
• To improve the performance of medical image classification and segmentation tasks using limited datasets.

Main Methods:

• Proposed CDA-GAN incorporates channel attention within a CycleGAN framework for cross-domain tumor image generation.
• Implemented a semi-supervised spatial attention strategy to guide pixel-level feature information for tumor generation.
• Integrated spectral normalization to stabilize training and prevent discriminator mode collapse.

Main Results:

• CDA-GAN generated diverse samples, expanding dataset scale and improving medical image diagnosis and segmentation.
• Outperformed state-of-the-art generative data augmentation on BraTS and TCIA datasets for classification and segmentation.
• Achieved significant improvements in accuracy, AUC, Recall, F1, Dice, Sensitivity, HD95, and mIOU.

Conclusions:

• CDA-GAN effectively enhances medical image datasets through cross-domain generation and attention mechanisms.
• The proposed model significantly improves diagnostic and segmentation accuracy, particularly in data-limited scenarios.
• CDA-GAN offers a valuable solution for improving medical image analysis tasks, including brain tumor classification and segmentation.