mertcaglar/segm repository overview

⁠🔄 WSI Downscaling

To downscale WSIs:

1. Run the dsconvert.py script.
2. Update folder paths in the script to match your local directory structure.
3. WSIs are split into patches (default: 2048x2048), and each patch is individually downscaled based on a specified scale factor.
4. Segmentation masks are generated from GeoJSON annotations provided with the dataset.
5. Downscaled patches are reassembled into .PNG images and masks using lossless compression.

⁠📁 Folder Structure

The expected folder structure is as follows:

ICIP2025/
├── geojson/
├── images/
├── dataset/
│   ├── images/
│   │   ├── train/
│   │   ├── test/
│   │   └── validation/
│   └── annotations/
│       ├── train/
│       └── validation/

⁠🐳 Docker Support

A prebuilt Docker image is available:

docker pull mertcaglar/segm

To run the container with GPU support and mount the current working directory:

docker run --gpus all -it -v $(pwd):/host -w /host mertcaglar/segm

This will launch the container and run the adaptive_trainer.py script.

⁠🧪 Experimentation & Training

⁠🧬 Adaptive Data Augmentation

All training experiments use adaptive data augmentation, optimized via the OpenAI API. To use this feature:

• Provide your API key in augmentation_strategy.py.
• Adjust the prompt to define your augmentation policy.
• adaptive_trainer.py will then apply the generated augmentations using the Albumentations⁠ library (Buslaev et al., 2020⁠).

⁠🛠️ Training Configurations

⁠🔍 Probability Matrix Inference

Once models are trained using the configurations above, generate probability maps for test images using:

infer_probs.py

Make sure to modify the script configuration to match your model and dataset settings.

⁠🧮 Top-N Soft Biased Voting

Use Top-N Soft Biased Voting to ensemble predictions across the best-performing models. To generate final masks from the probability matrices:

softvote_topN.py

This method improves accuracy by aggregating predictions from multiple models using soft-weighted voting.

⁠🧠 DPT Architecture Overview

DPT (Dense Prediction Transformer) is a vision transformer tailored for dense prediction tasks like semantic segmentation.

• Replaces conventional convolutional backbones with a transformer-based encoder, enabling global context at every layer.
• Intermediate transformer tokens are reconstructed into spatial feature maps and progressively decoded via a convolutional decoder.
• Offers fine-grained segmentation and superior global consistency compared to CNN-based approaches.