PathMNIST Image Classification

An experimental comparison of a classical machine-learning baseline and two neural-network architectures for nine-class histopathology image classification.

Portfolio note: This is a four-person university team project. I contributed to the Random Forest and MLP work, integrated the final experiment notebook, verified outputs against the report, and helped resolve implementation inconsistencies. The repository is a curated, public-facing version of the team submission; it is not presented as a solo project.

Highlights

• Task: classify 28 × 28 × 3 PathMNIST histopathology image patches into nine tissue classes.
• Models: Random Forest baseline, fully connected MLP, and CNN.
• Evaluation: stratified validation for hyperparameter search; held-out test accuracy, macro F1, class-wise metrics, and confusion matrices.
• Best model: CNN with 84.16% test accuracy and 0.8363 macro F1 on 8,000 held-out examples.

Results

Model	Input representation	Test accuracy	Macro F1	Training time
CNN	Normalised RGB images	0.8416	0.8363	68.39 s
Random Forest	Flattened grayscale pixels	0.5251	0.5023	4.77 s
MLP	Flattened normalised RGB pixels	0.5038	0.4527	16.63 s

The CNN's advantage is consistent with its spatial inductive bias: convolutional filters preserve local texture information, whereas the Random Forest and MLP work from flattened pixels.

Dataset

The experiment uses the course-provided subset of PathMNIST, a 2D histopathology image dataset from MedMNIST v2. The source data are not included in this repository. The original data are subject to the relevant dataset licence and course-distribution terms.

Classes: adipose, background, debris, lymphocytes, mucus, smooth muscle, normal colon mucosa, cancer-associated stroma, and colorectal adenocarcinoma epithelium.

Approach

1. Load and inspect image tensors and class distributions.
2. Create model-specific inputs: normalised RGB tensors for neural networks and flattened grayscale pixels for the random forest.
3. Use a stratified validation split for lightweight hyperparameter searches.
4. Train selected final models independently from the tuning cells.
5. Compare accuracy, macro F1, per-class reports, confusion matrices, and training time on the test set.

The final CNN contains three convolution-plus-max-pooling blocks with 32, 64, and 128 filters, followed by a 128-unit dense layer and dropout. It was trained with Adam and sparse categorical cross-entropy.

Reproducing the notebook

1. Create a Python environment and install dependencies:

   pip install -r requirements.txt

2. Place the authorised dataset files in Assignment2Data/ beside the notebook:

   Assignment2Data/
   ├── X_train.npy
   ├── X_test.npy
   ├── y_train.npy
   └── y_test.npy
   

3. Start Jupyter and run notebooks/pathmnist_model_comparison.ipynb from top to bottom.

The tuning cells can take appreciably longer than the final-model cells. Their outputs are retained for inspection.

Continuous verification

GitHub Actions runs a fast model smoke test on every push. It installs the project runtime and builds, trains for one epoch, and predicts with the Random Forest, MLP, and CNN on synthetic 28 × 28 RGB inputs. This validates the environment and end-to-end model paths without redistributing the course-provided dataset or misrepresenting synthetic-test results as the reported experiment metrics.

Tech stack

Python · TensorFlow/Keras · scikit-learn · NumPy · pandas · Matplotlib · Seaborn

Attribution

Dataset: J. Yang et al., MedMNIST v2: A large-scale lightweight benchmark for 2D and 3D biomedical image classification, Scientific Data, 2023. See MedMNIST for dataset details and licensing.

Please do not reuse this repository for assessed work. It is shared as a portfolio record and learning reference.