TransMamm

Transformer-Based TransMamm: A Density-Aware Hybrid Vision Transformer Framework for Robust Lesion Detection in Dense Breast Tissue

TransMamm is an ongoing research project focused on developing a deep learning framework for robust mammographic image analysis, with particular emphasis on lesion detection in dense breast tissue.

The project explores a hybrid architecture combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to capture both local visual features and global contextual information from mammographic images.

---

Research Motivation

Dense breast tissue can make lesions difficult to distinguish from normal fibroglandular tissue, because both may appear with similar intensity on mammograms.

This project investigates whether a density-aware hybrid CNN–Vision Transformer framework can improve lesion representation and robustness in challenging mammographic imaging scenarios.

---

Research Areas

• Machine Learning
• Deep Learning
• Computer Vision
• Medical Image Analysis
• Mammographic Image Analysis
• Breast Lesion Detection
• Convolutional Neural Networks (CNNs)
• Vision Transformers (ViTs)
• Density-Aware Attention
• Self-Supervised Learning
• Contrastive Learning
• Domain Generalization
• Explainable AI (XAI)
• Score-CAM

---

Datasets

The project explores publicly available mammographic datasets, including:

Dataset	Role
CBIS-DDSM	Primary dataset for initial development and experimentation
INbreast	External validation
CMMD	External validation
MIAS	External validation

Note: The datasets are not included in this repository due to their size and licensing considerations.

---

Planned Architecture

The proposed TransMamm framework is based on a hybrid CNN and Vision Transformer architecture:

                 Mammographic Image
                         │
              ┌──────────┴──────────┐
              │                     │
              ▼                     ▼
             CNN                   ViT
              │                     │
              └──────────┬──────────┘
                         │
                         ▼
              Density-Aware Fusion
                         │
                         ▼
                Feature Representation
                         │
                         ▼
                 Lesion Detection
                         │
                         ▼
                  Classification
                         │
                         ▼
                    Score-CAM

---

Project Structure

TransMamm/
│
├── notebooks/
│   └── CBIS_DDSM_Analysis.ipynb
│
├── src/
│   ├── preprocessing.py
│   ├── dataset.py
│   ├── model.py
│   ├── train.py
│   └── evaluate.py
│
├── results/
│
├── README.md
└── requirements.txt

The project structure will be updated as the research implementation progresses.

---

Current Progress

• CBIS-DDSM dataset exploration
• Image-path mapping
• Dataset preprocessing
• Patient-level data splitting
• Mammogram preprocessing pipeline
• CNN feature extraction
• Vision Transformer integration
• Density-aware attention
• Hybrid CNN–ViT architecture
• Model training
• Model evaluation
• Cross-dataset validation
• Score-CAM explainability
• Ablation studies

---

Technologies

• Python
• PyTorch
• torchvision
• NumPy
• Pandas
• OpenCV
• scikit-learn
• Matplotlib
• Jupyter Notebook
• kaggle GPU

---

The main objectives of this project are to:

• Develop a hybrid CNN–Vision Transformer framework for mammographic image analysis.
• Incorporate breast density information into the learning process.
• Improve lesion representation in dense breast tissue.
• Investigate model robustness across different mammographic datasets.
• Explore explainable AI techniques for interpreting model predictions.

---

Project Status

🚧 Research in Progress

This repository is actively being developed as part of an undergraduate research project. Implementation details, experiments, evaluation results, and findings will be added as the research progresses.

---

Author

Mayeda Islam
Central University of Science and Technology (CUST)
Computer Science and Engineering

Research Interests: Machine Learning · Deep Learning · Computer Vision · Medical Imaging · Vision Transformers · Explainable AI

---

Disclaimer

This project is intended for academic and research purposes only. The models developed in this repository are not intended for clinical diagnosis or medical decision-making.