MammoDenseSegNet: A Context-Aware Deep Learning Model for Dense Tissue Segmentation in Digital Mammograms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate breast density quantification is crucial for breast cancer risk assessment.
• Segmenting dense breast tissue in mammograms is challenging due to variations in appearance and imaging devices.
• Existing methods struggle with accuracy, particularly for low-density tissues.

Purpose of the Study:

• To develop and evaluate MammoDenseSegNet, a novel deep convolutional neural network for enhanced segmentation of dense breast tissue.
• To improve the accuracy and robustness of breast density quantification in mammograms.

Main Methods:

• MammoDenseSegNet employs an encoder-decoder architecture with an adaptive dual attention module and a multi-kernel receptive field module.
• A multi-scale dice loss with deep supervision was utilized to enhance learning across decoder levels.
• The model was evaluated on two public (VinDR-Mammo, EMBED) and one private mammogram dataset (1499 images).

Main Results:

• MammoDenseSegNet achieved high performance across diverse conditions (Recall: 0.64-0.90, Dice: 0.63-0.91).
• The model significantly outperformed a VGG16-based state-of-the-art algorithm (p < 0.001).
• Significant improvements were observed for low-density tissues, where MammoDenseSegNet demonstrated clinical utility (Recall: 0.66, Dice: 0.63) compared to the baseline's failure (Recall: 0.14, Dice: 0.16).

Conclusions:

• MammoDenseSegNet offers a robust and accurate solution for dense breast tissue segmentation in mammography.
• The proposed deep learning approach enhances breast cancer risk assessment capabilities.
• MammoDenseSegNet shows particular promise in improving the analysis of low-density breast tissues.