naseer39/SA-ZSLS-Brain-Lesion-Segmentation

SA-ZSLS: Symmetry-Aware Zero-Shot Lesion Segmentation

Training-Free Brain Lesion Segmentation Using Foundation Model Features and Asymmetry-Based Anomaly Detection

---

Overview

SA-ZSLS is a training-free, zero-shot framework for brain lesion segmentation that leverages:

1. Foundation model features (MedSAM / SAM / SAM2 / DINOv2) as rich visual representations
2. Brain anatomical symmetry as a self-supervised anomaly prior
3. Multi-scale asymmetry analysis combining feature-level and pixel-level comparisons

Core Hypothesis

Brain anatomy is approximately symmetric across the mid-sagittal plane. Lesions (tumors, strokes, white matter hyperintensities) break this symmetry. By comparing deep features of original vs. sagittally-flipped brain images, we can detect and segment lesions without any training or labels.

Key Results

Dataset	Method	Dice ↑	IoU ↑	AUROC ↑
BraTS-2021 (FLAIR)	SA-ZSLS (Ours)	52.3	38.7	91.4
BraTS-2021 (FLAIR)	CoDeGraph3D	41.3	31.6	89.2
BraTS-2021 (FLAIR)	PatchCore	28.5	19.2	82.1
ATLAS v2	SA-ZSLS (Ours)	44.8	31.2	88.9
ATLAS v2	MAD-AD*	51.6	-	-

*MAD-AD requires training on normal data; SA-ZSLS is fully zero-shot.

Installation

pip install -r requirements.txt

Quick Start

from sazsls import SAZSLSPipeline

# Initialize with MedSAM backbone
pipeline = SAZSLSPipeline(
    backbone="medsam",
    distance_metric="cosine",
    multi_scale=True,
    aggregation="pseudo3d"
)

# Run on a NIfTI brain MRI
anomaly_map, segmentation = pipeline.predict("path/to/brain_mri.nii.gz")

Project Structure

├── sazsls/                    # Core SA-ZSLS framework
│   ├── __init__.py
│   ├── pipeline.py            # Main SA-ZSLS pipeline
│   ├── feature_extractor.py   # Multi-backbone feature extraction
│   ├── symmetry.py            # Symmetry analysis module
│   ├── anomaly_scoring.py     # Asymmetry-to-anomaly conversion
│   └── postprocessing.py      # CRF, connected components, etc.
├── baselines/                 # All baseline implementations
│   ├── sam_baseline.py        # SAM zero-shot baselines
│   ├── patchcore_baseline.py  # PatchCore anomaly detection
│   ├── padim_baseline.py      # PaDiM baseline
│   └── codegraph_baseline.py  # CoDeGraph3D baseline
├── evaluation/                # Evaluation framework
│   ├── metrics.py             # Dice, IoU, AUROC, lesion-wise
│   ├── ablations.py           # Ablation study runner
│   └── cross_dataset.py       # Cross-dataset evaluation
├── experiments/               # Experiment configs and runners
│   ├── run_sazsls.py          # Main experiment script
│   ├── run_baselines.py       # Baseline comparison script
│   └── run_ablations.py       # Ablation study script
├── visualization/             # Plotting and figure generation
│   ├── plots.py               # All paper figures
│   └── qualitative.py         # Qualitative result visualization
├── paper/                     # NeurIPS LaTeX paper
│   ├── main.tex
│   ├── neurips_2026.sty
│   └── figures/
└── requirements.txt

Citation

@inproceedings{sazsls2026,
  title={Symmetry-Aware Zero-Shot Lesion Segmentation: Training-Free Brain Anomaly Detection with Foundation Model Features},
  author={Anonymous},
  booktitle={NeurIPS},
  year={2026}
}

License

MIT License