Convolutional Networks and Transformers for Mammography Classification: An Experimental Study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) are widely used for mammography analysis due to their feature extraction capabilities.
• Vision Transformers are emerging as a competitive alternative in medical imaging, sometimes surpassing CNN performance.

Purpose of the Study:

• To conduct an extensive experimental comparison of recent CNN and Vision Transformer architectures for whole mammogram classification.
• To evaluate model performance across various image resolutions and individual lesion categories.

Main Methods:

• Trained and tested 33 models (19 CNN-based, 14 Transformer-based) on the OMI-DB mammography dataset.
• Analyzed performance at eight different image resolutions.
• Evaluated classification for isolated lesion categories: masses, calcifications, focal asymmetries, and architectural distortions.

Main Results:

• Vision Transformers performed comparably to traditional CNNs like ResNet.
• Modern CNNs, specifically EfficientNet, demonstrated superior performance over Vision Transformers and other CNNs.
• Performance varied across different image resolutions and lesion types.

Conclusions:

• Modern CNN architectures, such as EfficientNet, show superiority for mammogram classification compared to Vision Transformers.
• Vision Transformers hold potential but require further development to consistently outperform advanced CNNs in this domain.
• The choice of architecture and image resolution impacts classification accuracy for specific mammographic features.