MammoViT: A Custom Vision Transformer Architecture for Accurate BIRADS Classification in Mammogram Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Mammography interpretation for breast cancer screening is vital but faces challenges like subtle features and reader variability.
• Existing computer-aided detection systems struggle with complex feature extraction and contextual understanding.
• The Breast Imaging-Reporting and Data System (BIRADS) classification is essential but difficult to automate accurately.

Purpose of the Study:

• To develop MammoViT, a novel hybrid deep learning framework combining ResNet50 and Vision Transformer for mammogram classification.
• To address the limitations of traditional methods in handling complex features and long-range dependencies in mammographic images.
• To improve the accuracy and reliability of automated breast cancer screening tools.

Main Methods:

• A multi-stage approach using ResNet50 for feature extraction and Vision Transformer for capturing spatial dependencies.
• Application of Synthetic Minority Over-sampling Technique (SMOTE) to handle class imbalance in the BIRADS dataset.
• Optimization via Keras Tuner and 5-fold cross-validation with early stopping for robust model training.

Main Results:

• MammoViT achieved 97.4% accuracy in classifying mammograms across different BIRADS categories.
• The model demonstrated effectiveness through comprehensive evaluation metrics, including classification reports and confusion matrices.
• Performance was validated against existing studies, showing superior or comparable results.

Conclusions:

• MammoViT successfully integrates ResNet50 and Vision Transformer architectures for medical image analysis.
• The framework effectively addresses data imbalance issues common in medical imaging datasets.
• MammoViT shows significant potential as a reliable tool to support clinical decision-making in breast cancer screening.