PanNuke Segmentation Playground

The Project

Nuclei segmentation is a basic task in computational pathology. This project covers the full training and evaluation pipeline for the PanNuke dataset, which contains more than 7,000 annotated image patches from 19 tissue types.

It is meant for students learning computer vision and anyone who wants to inspect the dataset or model outputs. The repository includes preprocessing, readable PyTorch implementations, training code, and an inference app.

Models

The repository includes four segmentation models:

• UNet: The original architecture from Ronneberger et al.
• UNet++: A nested U-Net that combines features across finer scales.
• ResNet/UNet (scratch): A ResNet-style encoder with a UNet decoder, trained from scratch.
• ResNet/UNet (pretrained): A pretrained ResNet encoder paired with the same custom decoder.

Interactive Web Application

The Streamlit app lets you:

• Browse the PanNuke dataset.
• Select a trained model.
• Run inference and overlay the predicted mask on the source image.

Code & Resources

The source code, training pipeline, and web app are on GitHub.

View on GitHub