Learning from adversarial medical images for X-ray breast mass segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Scarcity of labeled medical data hinders deep learning applications in medical imaging.
• Generating precise segmentation masks for medical images remains a significant challenge.
• Existing methods primarily focus on image classification and detection, not segmentation.

Purpose of the Study:

• To develop a method for generating realistic medical images with precise masks for improved lesion segmentation.
• To address the challenge of limited labeled data in mammogram analysis.
• To enhance the performance of deep learning models in breast mass segmentation.

Main Methods:

• A conditional generative adversarial network (cGAN) was developed to learn image-mask distributions.
• Adversarial lesion images with precise masks were generated using the cGAN.
• Generated images were combined with original data to augment the dataset.
• An improved U-net model was trained on the augmented dataset for segmentation.

Main Results:

• Experiments were conducted on the INbreast and a private hospital database.
• The proposed method achieved up to a 7% improvement in Jaccard index.
• Performance gains were observed compared to models trained solely on real images.

Conclusions:

• The study presents a novel approach for generating realistic X-ray breast mass images with masks.
• This method represents a significant step towards precise lesion segmentation in mammography.
• The findings support the utility of synthetic data generation for improving medical image analysis.