This project focuses on building an automated classification system to distinguish between malignant and benign breast tumors using ultrasound images. The system leverages deep learning techniques and optimization strategies to enhance model performance and support early diagnosis.
Breast cancer remains a global health issue, where early and accurate diagnosis is key. This project utilizes ResNet-based architectures to classify breast ultrasound images. Various strategies such as hyperparameter tuning, architectural modifications, and data transformations were applied to improve performance beyond baseline results.
- BreastMNIST: 28x28 grayscale breast ultrasound images
- Categories: Malignant vs. Benign
- Source: Breast Ultrasound Dataset
- Baseline Model: ResNet-18 (Pretrained on ImageNet)
- Improvements Made:
- Architectural modifications to ResNet18
- Custom loss functions
- Seed function for reproducibility
- Learning rate optimization using Optuna
- Hyperparameter tuning via grid search
| Metric | Baseline | Final Model |
|---|---|---|
| Accuracy (Test Set) | 0.863 | 0.8767 |
| AUC (Test Set) | 0.901 | 0.9105 |
| AUPR | — | 0.9238 |
| F1 Score | — | 0.8821 |
| 5-Fold CV Accuracy | — | 0.8040 |
| 5-Fold CV AUC | — | 0.8123 |
- AUC vs Accuracy: AUC reflects model’s ranking ability; higher AUC indicates better class separation.
- AUPR vs F1 Score: AUPR balances precision and recall across thresholds; F1 measures harmonic mean.
- Cross-validation: Used for robust performance estimation and to reduce variance from single test splits.
- Early approaches relied on feature engineering + SVMs.
- CNNs like ResNet have outperformed traditional methods.
- ResNet18 showed better performance than ResNet50 on BreastMNIST in prior studies.
- Python
- PyTorch
- Optuna
- Scikit-learn
- Matplotlib/Seaborn for evaluation
- Deep Residual Learning (He et al., 2016)
- BreastMNIST Dataset - Nature
- Ultrasound Classification Research (2022)
This work was carried out as part of an independent research project aimed at applying AI in healthcare to support early cancer detection through improved diagnostic models.
Feel free to contribute, report issues, or fork this repo to build upon it.