Project: PanNuke Segmentation Playground

The Goal: Education and Exploration

Nuclei segmentation in histopathology images is a foundational task in computational pathology. This project was created to be both a powerful tool and an open educational resource. It provides a complete ecosystem for training and evaluating deep learning models on the PanNuke dataset, which contains over 7,000 annotated image patches from 19 different tissue types.

Whether you're a student exploring computer vision and seeking PyTorch implementations, or a medical practitioner interested in interacting with this dataset, this repository offers a transparent and hands-on approach. Its core philosophy is to demystify these models by providing everything from the data preprocessing pipeline to an interactive web application for inference.

Architectures Implemented from Scratch

A key feature of this project is the collection of well-known segmentation architectures, all coded from the ground up in PyTorch to maximize understanding and customizability. The available models include:

UNet: The classic architecture from Ronneberger et al., a staple in biomedical image segmentation.
UNet++: An advanced, nested U-Net design that captures features at finer-grained scales for more precise segmentations.
ResNet/UNet (Scratch): A powerful hybrid combining a ResNet-style encoder with a UNet decoder, built entirely from scratch.
ResNet/UNet (Pretrained): Leverages a pretrained ResNet encoder for transfer learning, combined with the custom UNet decoder for faster, more robust training.

Interactive Web Application

To make model exploration intuitive and accessible, the project includes a web application built with Streamlit. This interface allows you to:

Browse through the PanNuke dataset on the fly.
Select any of the trained models.
Instantly perform inference and visualize the predicted segmentation mask overlaid on the original image.

This hands-on tool bridges the gap between training a model and seeing its tangible results, making it ideal for demonstrations, debugging, and educational purposes.

Code & Resources

The complete source code, training pipeline, and web application are available on GitHub. Dive in to train your own models or explore the pre-built implementations!

View on GitHub