kiselyovd/brain-mri-segmentation

Brain MRI Segmentation

A SegFormer-B2 model for FLAIR-abnormality masks on 2D LGG MRI slices. Native safetensors weights, an explicit foreground evaluation contract and original-source split identities are available in the source repository.

Current artifact and evaluation

Replayed on the reconstructed original patient split · 387 slices / 11 patients

Metric	Value
Foreground Dice	88.08%
Foreground IoU	78.70%

Foreground scores pool confusion counts across all test pixels. Historical checkpoint training membership is unverified; this is not an independently established held-out benchmark or a replay of the missing U-Net baseline. Patient-macro Dice: see the full report.

Evaluation record · Pinned model

The result is a new replay with recorded model and data hashes. The historical checkpoint's training membership is unverified. The original hand-rolled U-Net checkpoint has not been recovered, so there is no corrected historical main-versus-baseline comparison. Legacy reports retain the old aggregates separately.

Run the model

Use Transformers 5.5.4 or the project's locked environment, including torchvision.

import numpy as np
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForSemanticSegmentation

REPO = "kiselyovd/brain-mri-segmentation"
REVISION = "48c79adfde8b063e74737357ab20a6f7f67a0b1b"
processor = AutoImageProcessor.from_pretrained(
    REPO, revision=REVISION, backend="torchvision", trust_remote_code=False
)
model = AutoModelForSemanticSegmentation.from_pretrained(
    REPO, revision=REVISION, use_safetensors=True, trust_remote_code=False
).eval()
image = Image.open("slice.tif").convert("RGB")
# Float32 RGB [0,255] avoids uint8 rounding during tensor resizing.
inputs = processor(np.asarray(image, dtype=np.float32), return_tensors="pt")
with torch.inference_mode():
    logits = model(**inputs).logits
    logits = torch.nn.functional.interpolate(
        logits, size=(256, 256), mode="bilinear", align_corners=False
    )
    mask = logits.argmax(dim=1)[0]

The processor uses 256² inputs, rescaling 1/255, no ImageNet normalization and no label reduction. Actual predictions matched the repository path on eight committed TIF samples and one seeded nonsquare image: zero disagreeing mask pixels. Numerical parity receipt. This verifies computation paths rather than quality on unseen patients. The output mask corresponds to the resized input.

Original data and protocol

Mateusz Buda's LGG MRI dataset contains 110 TCGA-LGG patients and 3929 image/mask pairs. The three stored image channels represent MRI sequences; missing sequences are substituted as documented by the dataset author. A slice's binary mask labels FLAIR abnormality.

The original source archive was retrieved and hashed. Running the sorted-patient/seed 42 preparation recipe recovers 3133/409/387 slices and 88/11/11 patients. Source receipt · Complete split manifest · Evaluation contract.

Foreground Dice/IoU pool TP/FP/FN over all test pixels. Empty prediction and target score 1; false positives on an empty target score 0. The full report contains patient-macro measures, patient-cluster bootstrap intervals, per-slice/patient counts and prediction/target arrays. The limited cohort, 2D input and unresolved historical training lineage limit interpretation; clinical validation has not been performed.

Component licenses and attribution

Original project code is MIT. Dataset material and the source-derived manifest carry CC-BY-NC-SA-4.0 terms and attribution to Mateusz Buda and TCIA/TCGA-LGG. The upstream SegFormer card labels its license other and links NVIDIA's research/evaluation license; no independent unrestricted-MIT grant is asserted for the fine-tuned weights. Component details · Software citation.