LightweightUNet: Multimodal Deep Learning with GAN-Augmented Imaging Data for Efficient Breast Cancer Detection.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Breast cancer is a leading global cancer, necessitating early, automated, and precise detection methods.
• Existing AI techniques for breast cancer detection often involve high computational costs and complexity.

Purpose of the Study:

• To introduce an innovative LightweightUNet hybrid deep learning (DL) classifier for accurate breast cancer classification.
• To develop a DL model with a low computational cost and adaptive capabilities for efficient breast cancer detection.

Main Methods:

• A multimodal approach using 13,000 images from mammogram imaging (MGI) and ultrasound imaging (USI) from seven diverse sources.
• Data preprocessing included resizing to 256x256 pixels and normalization using Box-Cox transformation.
• Generative Adversarial Network (GAN) model StyleGAN3 was used to generate 10,000 synthetic ultrasound images to augment the smaller USI dataset. Experiments were conducted on real and real + GAN-augmented datasets using 5-fold cross-validation.

Main Results:

• The LightweightUNet model achieved 86.87% accuracy on the real dataset without augmentation.
• Performance significantly improved on the real + GAN-augmented dataset, reaching 96.35% accuracy.
• The multimodal approach using LightweightUNet demonstrated substantial performance enhancements, including a 9.48% increase in accuracy.

Conclusions:

• The proposed LightweightUNet model offers an effective solution for breast cancer classification with reduced computational demands.
• The integration of multimodal imaging and GAN-based data augmentation substantially boosts classification performance.
• The model shows significant potential for practical application in clinical settings for improved breast cancer diagnosis.