Breast Cancer Model Comparison
Comparative mammography and histopathology experiments across convolutional, attention and transformer architectures.
2025
Kaggle datasets · free Colab training · model comparisons.
Comparative imaging experiments
Breast Cancer Model Comparison studies mammography and histopathology across approximately three to four Kaggle datasets. Training used available free Google Colab accelerators. Evaluation metrics were computed and visualized across convolutional, attention-augmented and transformer configurations.
Mammograms and histopathology images represent different modalities. Preparation and comparison remain attached to the corresponding task rather than one interchangeable image collection.
Architecture comparison
Scroll horizontally to compare.
| Experiment | Model family | Comparison focus |
|---|---|---|
| EfficientNetV2 + CBAM | Attention-augmented CNN | Channel and spatial attention within a convolutional feature pipeline. |
| CNN / ViT hybrid | Hybrid experiment | Combining convolutional representation with transformer-based processing. |
| ViT variants | Vision transformers | Patch/token representations and attention-based image modeling. |
| ConvMixer | Patch-based convolutional architecture | Spatial and channel mixing through convolutional operations. |
| MobileNetV4 | Efficient convolutional family | An efficiency-oriented architecture in the comparison. |
| ResNet-50 | Residual convolutional network | A residual-network reference alongside newer model families. |
Training and comparison
- Dataset / modality
- Image preparation
- Model training
- Evaluation metrics
- Visual comparison
Results are compared within the corresponding dataset and experiment context across the model families.
Engineering decisions
Dataset preparation precedes architecture configuration, training and evaluation. CBAM adds channel/spatial attention to EfficientNetV2 convolutional features; this is distinct from a CNN–ViT hybrid. ViT variants and hybrid configurations explore token representations/global context, while ConvMixer uses patch-based convolutional mixing. MobileNetV4 and ResNet-50 contribute mobile-oriented and residual convolutional alternatives.
Comparative visualizations make the architecture differences and evaluation context readable across runs. The workflow centers on preparing image inputs, configuring different inductive biases, accelerator-backed training and interpreting model output. Each comparison retains its modality, dataset and training context alongside the evaluation.