Hybrid transformer-based model for mammogram classification by integrating prior and current images.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Systems
• Oncology and Radiology Research

Background:

• Mammography is vital for early breast cancer detection, but manual image analysis is challenging.
• Current methods face limitations in efficiency and accuracy for computer-aided diagnosis.
• Radiologists' detailed examination process requires significant time and expertise.

Purpose of the Study:

• To introduce a CNN-Transformer model for breast cancer classification using mammographic analysis.
• To leverage temporal changes by analyzing both prior and current mammograms.
• To enhance the efficiency and accuracy of computer-aided diagnosis systems.

Main Methods:

• Developed a CNN-Transformer model integrating position-wise feedforward networks and multi-head self-attention.
• Employed positional encoding and channel attention to highlight critical spatial features for tissue differentiation.
• Utilized focal loss (FL) to address difficult classifications, reducing false negatives and positives.

Main Results:

• The proposed model outperformed eight baseline models in accuracy (ACC), sensitivity (SEN), precision (PRE), specificity (SPE), F1 score, and AUC.
• Achieved high performance metrics: ACC 90.80%, SEN 90.80%, PRE 90.80%, SPE 90.88%, F1 90.95%, AUC 92.58%.
• Model code and information are publicly available.

Conclusions:

• The CNN-Transformer model effectively integrates prior and current images, reducing long-range dependencies for nuanced classification.
• Focal loss application significantly improved sensitivity and specificity by reducing false positive and negative rates.
• The model demonstrated low error rates for various abnormalities, showing potential for reliable early breast cancer detection.