Connected-UNets: a deep learning architecture for breast mass segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Radiologists analyze mammograms for breast cancer detection, identifying suspicious lesions and tumors.
• Artificial intelligence (AI) and deep learning offer automated breast mass segmentation to aid diagnosis.
• UNet and its variants are leading models for medical image segmentation, showing promise in mammography.

Purpose of the Study:

• To propose a novel deep learning architecture, Connected-UNets, for improved automatic breast mass segmentation.
• To integrate Atrous Spatial Pyramid Pooling (ASPP) within UNet architectures to enhance contextual information.
• To evaluate the proposed Connected-UNets architecture on standard and Attention/Residual UNet variations.

Main Methods:

• Developed Connected-UNets by linking two UNets with modified skip connections.
• Integrated ASPP into standard UNets, Attention UNets (AUNet), and Residual UNets (ResUNet).
• Utilized public (CBIS-DDSM, INbreast) and private datasets, augmented with synthetic data from CycleGAN.

Main Results:

• Achieved high Dice scores: 89.52% (CBIS-DDSM), 95.28% (INbreast), and 95.88% (private dataset).
• Obtained high Intersection over Union (IoU) scores: 80.02% (CBIS-DDSM), 91.03% (INbreast), and 92.27% (private dataset).
• Demonstrated superior automatic mass segmentation performance compared to existing methods.

Conclusions:

• The proposed Connected-UNets architecture significantly improves automatic breast mass segmentation accuracy.
• Integrating ASPP enhances the model's ability to capture contextual information for better segmentation.
• The findings support the use of advanced AI models like Connected-UNets in clinical mammography for improved breast cancer detection.