# Finding Cancer in Ultrasound Images
From pixel logistic regression to transfer learning
BreastMNIST ultrasound — metrics, trade-offs, and Grad-CAM explanations of what each model attends to
Not a medical device. This is an educational computer-vision case study on a public 128×128 benchmark. It is not validated for screening or diagnosis, and it should not be used to interpret clinical images.
## Project Overview
Radiologists reading breast ultrasound look for irregular hypoechoic masses, angular margins, and posterior acoustic shadowing. This project asks how far a sequence of machine-learning models can get on the same task — **malignant vs benign/normal** — when the images are the public [BreastMNIST](https://medmnist.com/) split of the BUSI ultrasound collection.
The modelling ladder is deliberate. A majority-class dummy sets the accuracy floor (never predicting cancer still scores ~73% because malignancy is the minority class). A linear model on PCA-reduced pixels is next, then classical computer-vision features (HOG and LBP) with a random forest, then a small CNN trained from scratch, and finally a frozen ImageNet ResNet-18 with a new classification head. Every model is scored on the **same official test set**, with decision thresholds chosen on the validation split.
The interesting part is not a single accuracy number. Precision, recall, specificity, ROC-AUC and PR-AUC move differently as the models get more expressive — and Grad-CAM overlays show whether the conv nets are looking at the mass or at a scan artefact.
## Project Structure
Five notebooks, from the images through to a held-out comparison dashboard:
1. Images & Labels
Explore →
Official splits, class balance, montages, and mean-image differences
2. Classical Models
Explore →
Dummy classifier, PCA + logistic regression, HOG/LBP random forest
3. Deep Learning
Explore →
A small CNN from scratch and a frozen ResNet-18 transfer head
4. Interpretation
Explore →
Pixel coefficient maps, texture importances, and Grad-CAM case grids
5. Evaluation
Explore →
Precision, recall, ROC/PR, calibration, and which metric to trust
## Key Questions
- If a model never predicts cancer, how high can accuracy still be — and why is that useless here?
- Do hand-crafted edge and texture features beat a linear model on flattened pixels?
- Does a CNN trained on 546 images outperform those classical features, or do we need transfer learning?
- When Grad-CAM lights up, is it the mass, the acoustic shadow, or a measurement crosshair?
## Preview Figures
## Dataset
- **Source:** [BreastMNIST](https://medmnist.com/) (MedMNIST v2/v3), derived from the BUSI breast-ultrasound collection
- **Task:** binary classification, **malignant = positive class** (MedMNIST's 0/1 labels are remapped in the notebooks)
- **Size:** 780 images at 128×128 grayscale — 546 train / 78 val / 156 test
- **Licence:** MedMNIST **CC BY 4.0**; cite Yang et al., *Scientific Data* (2023) and Al-Dhabyani et al., *Data in Brief* (2020)
- **Bundled file:** `data/breastmnist_128.npz` (no live download at book-build time)
## Technical Stack
**Python** • MedMNIST • scikit-learn • scikit-image • PyTorch • torchvision • Plotly • Grad-CAM
---