🧠 BrainSec2.0 — WSI Segmentation Toolkit

Figure 1: End-to-End Whole-Slide-Image (WSI) Segmentation Pipeline using SegFormer

Overview

BrainSec2.0 is a transformer-based segmentation toolkit for large-scale Whole-Slide Images (WSI) of the human brain. It extends SegFormer (Xie et al., 2021) with LoRA / QLoRA fine-tuning, quantization, and ONNX-based inference for GPU and CPU environments. The toolkit achieves research-grade accuracy while remaining lightweight enough for macOS laptops and low-resource desktops.

Key Features

Category	Description
Segmentation	White- and gray-matter segmentation on high-resolution WSIs
Fine-Tuning	Four LoRA / QLoRA configurations for flexible adaptation
Performance	Full-slide inference in ≈ 3 min (GPU) or ≈ 15 min (CPU / M-series Mac)
Deployment	Quantized ONNX export for efficient CPU inference
Platform Support	macOS, Linux, NVIDIA GPUs, and HPC clusters
Ease of Use	Plug-and-play scripts + reproducible Conda environment

Quickstart

Step 1 — Setup Environment

Clone the repository and create the Conda environment:

conda env create -f configs/server_env.yml -n brainsec2
conda activate brainsec2

Step 2 — Download Models

Model weights are large and hosted externally on Google Drive at: https://drive.google.com/drive/folders/1NbLP4E-m5RhgTmHj4mcee1ZR6OcevJIm?usp=sharing

More public WSI data available at:

1. https://zenodo.org/records/1470797 (.svs)
2. https://datadryad.org/dataset/doi:10.5061/dryad.7h44j107j (.czi)

mkdir -p models
unzip ~/Downloads/model.zip -d models/

Step 3 — Run Inference

GPU Inference (≈ 3 min on NVIDIA GPU)

python src/inference/inference_main.py \
  --model-type pretrained \
  --model-dir models/segformer_pretrained \
  --wsi-path data/sample.svs \
  --output output.png \
  --batch-size 8

CPU Inference with ONNX (≈ 15 min on M-series)

python src/onnxformatting/inference_onnx.py \
  --onnx-model models/segformer_lora.onnx \
  --wsi-path data/sample.svs \
  --output output_onnx.png \
  --executor-provider CPUExecutionProvider

Fine Tuning

Finetune the base segformer model using a strategy

python src/inference/inference_main.py \
  --model-type lora \
  --model-dir models/ft_models/finetuned_lora \
  --base-model-dir models/segformer_pretrained \
  --wsi-path data/sample.svs

Documentation Report

https://drive.google.com/file/d/17flD9lxXUXJsLXH2l05qet6YfuAPcc4E/view?usp=sharing