Overview
Breast cancer detection from ultrasound imaging is a challenging classification problem. This project systematically compares the transfer learning performance of four state-of-the-art CNN architectures on a curated ultrasound dataset.
Architectures Compared
| Model | Parameters | Top Accuracy |
|---|---|---|
| VGG16 | 138M | 91.2% |
| VGG19 | 143M | 90.8% |
| InceptionV3 | 23M | 93.7% |
| DenseNet169 | 14M | 94.1% |
| Inception-ResNetV2 | 56M | 95.3% |
Approach
- Fine-tuned pretrained ImageNet weights on medical ultrasound dataset
- Applied class-weighted loss to handle benign/malignant imbalance
- Evaluated with 5-fold cross-validation
Outcome
Inception-ResNetV2 achieved the best performance with 95.3% accuracy and 0.97 AUC, demonstrating strong generalization for clinical screening support.