Multi-view fusion-based local-global dynamic pyramid convolutional cross-tansformer network for density...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Breast density is a key breast cancer risk factor.
• Current mammography classification methods often fail to fully leverage multi-view data, limiting accuracy.
• Accurate breast density classification is crucial for risk stratification and early detection.

Purpose of the Study:

• To develop an advanced deep learning model for improved breast density classification using multi-view mammography.
• To enhance the utilization of both local and global features from mammographic images.
• To address class imbalance issues in public datasets for more robust model performance.

Main Methods:

• Proposed a local-global dynamic pyramidal-convolution transformer network (LG-DPTNet) for multi-view fusion.
• Employed a dynamic pyramid convolutional network for adaptive single-view feature extraction (local and global).
• Utilized a cross-transformer to integrate fine-grained and global contextual information across views.
• Implemented an asymmetric focal loss function to handle class imbalance during training.

Main Results:

• Achieved an Area Under the Curve (AUC) of 96.73% on the CBIS-DDSM dataset.
• Attained an AUC of 91.12% on the INbreast dataset.
• Demonstrated superior performance compared to baseline and state-of-the-art methods.

Conclusions:

• The proposed LG-DPTNet effectively utilizes multi-view mammography information for breast density classification.
• The model significantly outperforms existing methods in terms of classification accuracy.
• This approach offers a promising advancement for breast cancer risk assessment through improved mammographic analysis.