BUS-GAN

BUS-GAN applies semi-supervised Generative Adversarial Networks to segment breast lesions in ultrasound images by leveraging both annotated and unannotated data.

Key Features:

• Segmentation Base Network (BUS-S): Densely extracts multi-scale features to accommodate individual variance in breast lesions and improve robustness of tumor-region segmentation.
• Evaluation Base Network (BUS-E): Incorporates a dual-attentive-fusion block with two independent spatial-attention paths on predicted segmentation maps, integrating geometrical-level and intensity-level information from the original image to amplify lesion–background distinction and guide BUS-S.
• Adversarial Training: Employs adversarial training in which BUS-E directs BUS-S to produce segmentation maps that closely resemble ground-truth distributions, enhancing overall segmentation quality.

Scientific Applications:

• Clinical Practice Enhancement: Provides automated breast-lesion segmentation for computer-aided diagnostic systems, reducing reliance on manual annotation and mitigating subjective biases from varying physician experience.
• Utilization of Unannotated Data: Leverages unannotated breast ultrasound images within a semi-supervised framework to improve segmentation performance and reduce annotation requirements.

Methodology:

BUS-GAN was trained on 2,000 in-house breast ultrasound images (100 annotated, 1,900 unannotated) and evaluated on an in-house test set of 800 images and a public set of 163 images; comparative analysis against state-of-the-art semi-supervised segmentation methods assessed segmentation accuracy.