🧠 U-Net Paper Reimplementation — From Scratch and Faithful to the Original

Faithful implementation of the original U-Net architecture with weight maps, elastic deformation, and full training pipeline — based on the 2015 paper.

This project is a deep, from-scratch reimplementation of the original U-Net: Convolutional Networks for Biomedical Image Segmentation paper (Ronneberger et al., 2015).
The implementation is divided into beginner-friendly Kaggle notebooks, each dedicated to a core component of the original architecture and training setup — no shortcuts, no black boxes.

📌 This repository serves as a landing page to the full notebook series and highlights what makes this implementation faithful to the original work.

---

🔗 Kaggle Notebook Series

Part	Description	Link
Part 1	U-Net Architecture from Scratch	🔗 Notebook
Part 2	Elastic Deformation & Augmentation	🔗 Notebook
Part 3	Weight Map Generation	🔗 Notebook
Part 4	Dataset Preparation	🔗 Notebook
Part 5	Training Loop & Results	🔗 Notebook
Bonus	Elastic Deformation Explained in Depth	🔗 Notebook

📌 Each notebook is structured, well-commented, and designed to be accessible for learners and reproducible for practitioners.

---

💡 What Makes This Different?

• ✅ Full U-Net architecture built from scratch, exactly as described in the original paper
• ✅ Implements elastic deformation using custom displacement fields and Gaussian smoothing
• ✅ Includes pixel-wise weight map generation to emphasize borders and handle class imbalance
• ✅ Reproducible data preparation pipeline: train/validation split, augmentation, and TFRecord generation
• ✅ Trains on 25,000+ augmented images using Kaggle TPUs for massive speedup (TPU-VM v3-8)
• ✅ Covers learning rate scheduling, callbacks, and checkpointing in training
• ✅ Clean modular split: architecture · augmentation · weight maps · data prep · training
• ✅ Uses public biomedical datasets (e.g. ISBI-style microscopy)
• ✅ Beginner-friendly walkthroughs — with motivation, visualizations, and linked components

---

⚠️ Faithful to the Paper: What's Usually Done Wrong

Most public U-Net implementations deviate from the original 2015 paper in critical ways.
This project avoids those shortcuts and replicates the original design as closely as possible:

• ❌ Uses padding='same' in conv layers
  ✅ Uses valid convolutions with manual input mirroring and padding to preserve spatial structure
• ❌ Assumes auto-alignment of encoder–decoder paths
  ✅ Performs explicit cropping to align encoder feature maps during concatenation
• ❌ Skips weight maps entirely
  ✅ Includes pixel-wise weight map generation to emphasize border regions (as per the paper)
• ❌ Ignores elastic deformation
  ✅ Implements elastic deformation from scratch using scipy.ndimage, matching the paper's core augmentation strategy for biomedical images

📌 This implementation is not just functional — it’s pedagogically faithful to the original U-Net design. If you're comparing U-Net repos, this one aims to teach why each detail matters.

---

🛠 Technologies Used

• Python · TensorFlow · Keras · NumPy · SciPy (scipy.ndimage) · Matplotlib
• Kaggle Notebooks for training and TPU execution
• Open-source biomedical datasets (e.g., ISBI, Cell Nuclei)

---

📘 Original Paper

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox.
"U-Net: Convolutional Networks for Biomedical Image Segmentation."
arXiv preprint arXiv:1505.04597 (2015)

---

🏷 Topics

U-Net · Semantic Segmentation · Biomedical Imaging · Deep Learning · Paper Reimplementation · Elastic Deformation · Keras · Image Segmentation

---

📥 Want the Code Here?

This repo currently links to well-structured Kaggle notebooks.
A PyTorch/Keras codebase version may be added in the future. ⭐️ Watch the repo to get notified.

---