Pathology Classifier

🧬 Overview

This repo contains a PyTorch implementation of a 10-class pathology image classifier.
The dataset consists of histopathology patches sorted into folders by tissue type (400px resolution).

Classes:

• Acinar tissue
• Dysplastic epithelium
• Fibrosis
• Lymph Aggregates
• Necrosis
• Nerves
• Normal ductal epithelium
• Reactive
• Stroma
• Tumor

The goal is to recreate an earlier model in PyTorch, train it on Raj's dataset, and evaluate performance with per-class metrics + confusion matrix.

---

📂 Project structure

pathology-classifier/
├── data/                # raw data (NOT in git)
│   └── classification/  # 10 class folders
├── notebooks/           # experiments
│   └── explore_data.ipynb
├── src/                 # training & evaluation scripts
│   ├── train_classify.py
│   ├── evaluate.py
│   ├── split_data.py
│   ├── models/          # custom architectures
│   └── utils/           # helpers
├── outputs/             # trained weights, plots (gitignored)
├── config.json          # hyperparams
├── requirements.txt     # dependencies
└── README.md

---

⚙️ Setup

1. Clone the repo:

   git clone https://github.com/tdiprima/pathology-classifier.git
   cd pathology-classifier

2. Create a virtual env (optional but recommended):

   python -m venv .venv
   source .venv/bin/activate   # (Linux/macOS)
   .venv\Scripts\activate      # (Windows)

3. Install dependencies:

   pip install -r requirements.txt

4. Place dataset under:

   data/classification/400p-Acinar tissue/*.png
   data/classification/400p-Dysplastic epithelium/*.png
   ...
   

---

🚀 Training

Run:

python src/train_classify.py

• Automatically splits data into train/val (80/20).
• Saves best model to models/DecaResNet_v2.pth (PyTorch model)
• Exports ONNX model to models/DecaResNet_v2.onnx (ONNX model).
• TorchScript model => models/DecaResnet.pt

You can tweak hyperparams in config.json.

---

📊 Evaluation

After training, run:

python src/evaluate.py

• Prints per-class precision/recall/F1.
• Generates confusion_matrix.png in outputs/.

---

🧪 Experiments

Use notebooks/explore_data.ipynb for:

• Visual sanity checks of dataset
• Trying different augmentations
• Quick evaluation on small subsets

---

📌 Notes

• Dataset is too large for GitHub → keep it local or use DVC (Data Version Control) / symlinks if needed.
• All configs (paths, batch size, image size, epochs, LR) live in config.json.
• Outputs (weights, plots) are gitignored.