---

Annotsim: Self-Supervised Anomaly Segmentation via Diffusion Models with Dynamic Transformer UNet (WACV 2024) 🚀

This research paper introduces a novel self-supervised anomaly detection method for image segmentation. It employs a diffusion model utilising a newly developed four-dimensional simplex noise function (Tsimplex) for improved efficiency and sample quality, especially in higher-dimensional and coloured images. The core of the model is a Dynamic Transformer UNet (DTUNet), a modified Vision Transformer architecture designed to handle time and noise image patches as tokens. Extensive experiments across three datasets demonstrate significant performance improvements over existing generative-based anomaly detection methods, particularly in medical imaging. The source code is publicly available.

Features

• Custom Diffusion Models: An extended UNet with selective denoising capabilities.
• Dynamic Transformer Blocks: Enhancements for dynamic anomaly segmentation.
• Simplex Noise Integration: Supports 3D/4D noise generation for improved feature diversity.
• Comprehensive Evaluation Metrics: Tools to calculate precision, recall, Dice score, and more.
• Visualization Tools: Generates diffusion videos and detection outputs for interpretability.

---

Example Outputs

Diffusion Training Visualization

Distribution Visualization & Histogram Plot

Distribution Visualization	Histogram Plot

Qualitative Results

Qualitative Result 1	Qualitative Result 2

Octave Visualization

SSIM Plot

Time Complexity Plots

MRI Translation

---

Anomaly Detection Results

Simplex Noise Example

Anomaly Detection Results

Simplex Noise Example

---

Project Structure

Annotsim/
├── src/
│   ├── models/               # Model architectures (UNet, transformer blocks, etc.)
│   ├── utils/                # Helper functions (dataset loading, noise generation, etc.)
│   ├── scripts/              # Training and evaluation scripts
├── requirements.txt          # Python dependencies
├── setup.py                  # Installation script
├── README.md                 # Project documentation
└── .gitignore                # Ignored files and directories

---

Installation

Prerequisites

• Python 3.8 or later
• CUDA-enabled GPU (optional but recommended for training)

Steps

1. Clone the repository:

   git clone https://github.com/MAXNORM8650/Annotsim.git
   cd Annotsim

2. Install dependencies:

   pip install -r requirements.txt

3. Install the repository as a package:

   pip install -e .

---

Usage

Train a Diffusion Model

To train a diffusion model:

python src/scripts/diffusion_training_UVW.py --argN

Replace argN with your desired configuration file from "test_args" file.

Evaluate a Model

To evaluate a model:

python src/scripts/detection.py --argN

---

Datasets

This project uses two publicly available datasets:

1. BRATS2021: A dataset for brain tumor segmentation.
2. Pneumonia X-Ray: A dataset for chest X-ray anomaly detection.

For more details, refer to the datasets' official documentation.

---

Results

Diffusion Videos

Generated videos during training and evaluation are saved in:

outputs/diffusion-videos/

Detection Outputs

Detection results are saved in:

outputs/detection-images/

---

Contributors

• Komal Kumar: GitHub Profile

The project imported from:

• AnoDDPM
• Predictive Convolutional Attentive Block
• Guided Diffusion

---

Citation

@InProceedings{Kumar_2025_WACV,
    author    = {Kumar, Komal and Chakraborty, Snehashis and Mahapatra, Dwarikanath and Bozorgtabar, Behzad and Roy, Sudipta},
    title     = {Self-Supervised Anomaly Segmentation via Diffusion Models with Dynamic Transformer UNet},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
    month     = {February},
    year      = {2025},
    pages     = {7917-7927}
}