Histopathology Image Classification

A deep learning project for colorectal cancer tissue classification using H&E-stained histopathology image patches. This repository fine-tunes an ImageNet-pretrained EfficientNet-B0 model on the NCT-CRC-HE-100K dataset and evaluates it on the independent CRC-VAL-HE-7K test set.

Overview

This project performs patch-level classification of colorectal histology images into 9 tissue categories:

• ADI: adipose tissue
• BACK: background
• DEB: debris
• LYM: lymphocytes
• MUC: mucus
• MUS: smooth muscle
• NORM: normal colon mucosa
• STR: cancer-associated stroma
• TUM: colorectal adenocarcinoma epithelium

The goal is to build a compact and reproducible baseline for colorectal tissue recognition using transfer learning.

Dataset

The project uses the public colorectal cancer histology datasets by Kather et al.

Dataset	Usage	Size	Description
NCT-CRC-HE-100K	Training	100,000 patches	H&E-stained colorectal tissue image patches
CRC-VAL-HE-7K	Testing	7,180 patches	Independent external validation set

All images are 224 x 224 pixel H&E-stained tissue patches. The external validation set is patient-independent from the training set, making it useful for evaluating model generalization.

Method

The pipeline uses transfer learning with EfficientNet-B0:

1. Load EfficientNet-B0 with ImageNet-pretrained weights.
2. Replace the classifier head with a 9-class output layer.
3. Resize and normalize images using ImageNet statistics.
4. Train with cross-entropy loss.
5. Optimize using Adam.
6. Save the best model according to test accuracy.

Model Configuration

Component	Setting
Architecture	EfficientNet-B0
Pretraining	ImageNet
Input size	224 x 224
Number of classes	9
Optimizer	Adam
Learning rate	0.001
Loss function	Cross-Entropy Loss
Batch size	64
Epochs	10

Results

Metric	Value
Test Accuracy	96.76%
Epochs	10
Batch Size	64

The model achieves strong classification performance on the independent CRC-VAL-HE-7K test set, showing that EfficientNet-B0 transfer learning is an effective baseline for colorectal histopathology patch classification.

Repository Structure

histopathology-classification/
├── README.md
├── models.py
├── train.py
├── predict.py
└── requirements.txt

File	Description
models.py	Defines the EfficientNet-B0 classification model
train.py	Training and evaluation pipeline
predict.py	Inference script for single-image prediction
requirements.txt	Python dependencies

Installation

Clone the repository:

git clone https://github.com/Qiaoli-Li-Res/histopathology-classification.git
cd histopathology-classification

Create a Python environment:

conda create -n histopathology python=3.10 -y
conda activate histopathology

Install dependencies:

pip install -r requirements.txt

Usage

Train the model

python train.py

This trains EfficientNet-B0 on the training dataset and evaluates the model on the test set.

Run prediction

python predict.py

This script can be used to run inference on a histopathology image patch with the trained model.

Example Workflow

# 1. Install dependencies
pip install -r requirements.txt

# 2. Train the classifier
python train.py

# 3. Run prediction
python predict.py

Notes

• This project performs patch-level tissue classification, not whole-slide image diagnosis.
• The reported result is based on the CRC-VAL-HE-7K external test set.
• For clinical usage, additional validation, calibration, interpretability analysis, and expert pathology review are required.
• Accuracy alone may not fully reflect model reliability, especially when class imbalance or staining variation exists.

References

• Kather, J. N., Halama, N., & Marx, A. 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo, 2018. https://doi.org/10.5281/zenodo.1214456
• Tan, M., & Le, Q. V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. ICML, 2019.ng · EfficientNet · PyTorch