Related Experiment Video
Updated: Jun 20, 2026
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
42.8K
Mammography classification with multi-view deep learning techniques: Investigating graph and transformer-based
Francesco Manigrasso1, Rosario Milazzo1, Alessandro Sebastian Russo1
1Politecnico di Torino, Dipartimento di Automatica e Informatica, Corso Duca degli Abruzzi 24, 10129, Turin, Italy.
Medical Image Analysis
|September 8, 2024
Summary
Deep learning models show promise for screening mammography, but challenges remain. Transformer-based architectures perform best, but ensembles of diverse models offer the most robust breast cancer classification.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Deep learning (DL) offers potential for automated screening mammography assessment.
- Challenges include low cancer prevalence, high-resolution images, and multi-view data integration.
- Weakly-supervised learning on exam-level labels is constrained by dataset size and accuracy.
Purpose of the Study:
- To evaluate novel transformer-based and graph-based architectures for multi-view mammography.
- To compare these against state-of-the-art convolutional neural networks (CNNs).
- To assess performance and interpretability in a weakly-supervised setting.
Main Methods:
- Extensive evaluation of transformer (ViT) and graph-based architectures.
- Comparison with multi-view CNNs on the CSAW dataset.
- Weakly-supervised training using exam-level labels on a middle-scale dataset.
Main Results:
- Transformer-based architectures demonstrated superior performance.
- Different architectures exhibit complementary strengths and weaknesses.
- Ensembling diverse architectures yielded more accurate and robust results than single models.
Conclusions:
- Multi-view architectures show significant potential for breast cancer classification, even with moderate datasets.
- Transformer and graph-based models offer advantages in integrating mammographic views.
- Detecting small lesions remains challenging without pixel-level supervision or specialized networks.
