U-Net vs. UNet++ for Nuclei Segmentation (DSB2018)

This repository contains a comprehensive Jupyter Notebook that implements, trains, and compares two deep learning architectures, U-Net and UNet++ (Nested U-Net), for image segmentation. The project aims to identify and segment cell nuclei from microscopy images provided by the Data Science Bowl 2018 (DSB2018) Kaggle competition.

The entire workflow is contained in a single notebook, demonstrating:

• EDA: In-depth Exploratory Data Analysis of the image and mask dataset.
• Preprocessing: Loading, resizing, and merging individual masks into a single binary mask per image.
• Model Implementation: PyTorch implementations of both U-Net and UNet++ (with deep supervision) from scratch.
• Training: A complete training pipeline including a combined BCE + Dice Loss, Adam optimizer, and Early Stopping.
• Evaluation: A robust 70/10/20 Train/Validation/Test split for fair comparison using the Intersection over Union (IoU) metric.
• Visualization: Detailed plots of training history and visual comparison of segmentation results.

---

🏛️ Model Architectures

This project implements and compares two of the most influential architectures for semantic segmentation.

1. U-Net

The foundational encoder-decoder architecture. It introduces "skip connections" that merge deep, semantic feature maps from the encoder with shallow, high-resolution feature maps in the decoder, enabling precise localization.

2. UNet++ (Nested U-Net)

An evolution of U-Net, it features "nested and dense skip connections" (as shown below) to bridge the semantic gap between the encoder and decoder. This project implements the deep supervision (L1-L4) variant, as described in the original paper, which helps the model produce accurate segmentation maps at multiple resolutions.

---

🚀 Final Results

Both models were trained under identical conditions (learning rate, batch size, loss function, etc.) until validation IoU stopped improving (Early Stopping).

Quantitative Results

The final performance was evaluated on the held-out 20% test set. UNet++ (with deep supervision) demonstrated a clear improvement in segmentation accuracy over the baseline U-Net.

Model	Test Loss (BCE+Dice)	Test IoU (Jaccard)
U-Net	0.1909	0.8476
UNet++ (Deep Supervision)	0.1348	0.8531

Training History

The training curves below illustrate the convergence behavior of both models.

Qualitative Comparison

Visual inspection of the test set predictions shows that UNet++ often produces cleaner and more accurate boundaries for overlapping nuclei.

---

📁 Project Structure

Nuclei-Segmentation-UNet-vs-UNetPlusPlus/
│
├── .gitignore          # Git ignore file
├── README.md           # This file
├── requirements.txt    # Python dependencies
│
├── notebooks/
│   └── unet_vs_unetpp.ipynb  # The main Jupyter Notebook with all code
│
└── assets/
    ├── unet_architecture.png   # (Ảnh bạn thêm vào)
    ├── architecture.png    # UNet++ architecture diagram
    ├── results_plot.png    # Screenshot of training history
    └── segmentation_demo.png # Screenshot of visual results

---

🛠️ How to Run

1. On Kaggle (Recommended)

The easiest way to run this project is to use the Kaggle environment.

1. Create a new Kaggle Notebook.
2. Upload the notebooks/unet_vs_unetpp.ipynb file.
3. Click "+ Add Data" and search for and add the "Data Science Bowl 2018" dataset.
4. Ensure the notebook's accelerator is set to GPU.
5. Run all cells.

2. Local Setup

You can also run the project on your local machine if you have a CUDA-compatible GPU.

1. Clone the repository:

   git clone [https://github.com/](https://github.com/)[YourUsername]/[YourRepoName].git
   cd [YourRepoName]

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

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

4. Download the data:

   • Download the DSB2018 dataset from Kaggle.
   • Extract the stage1_train/ folder.
   • Important: Open the notebook and modify the TRAIN_DIR variable (in the config cell) to point to the stage1_train/ directory on your local machine.

5. Run Jupyter:

   jupyter notebook notebooks/unet_vs_unetpp.ipynb

---

📦 Requirements

All required libraries are listed in requirements.txt. The primary libraries are:

• torch
• numpy
• matplotlib
• scikit-image
• tqdm
• jupyter