mhist-vote-count-deit-s16

MHIST Pathologist Vote-Count DeiT-S/16

This public repository documents the model evaluated in Pathologist Vote-Count Regression with a Rotation-Consistent Transformer for MHIST by Hanyue Shen and Xinling Liao (equal contribution), YH Intelligence Technology. It provides the architecture, inference interface, configuration, aggregate metrics, and citation material.

Correspondence: [email protected], [email protected].

Weight-release status

No trained weights are included in this public repository. The five checkpoints are retained in a separate private repository pending specific written permission from the MHIST data publisher for public redistribution of trained weights. The MHIST Research Use Agreement restricts redistribution and derivative works. See USE_RESTRICTIONS.md.

Model

• Backbone: DeiT-S/16 (deit_small_patch16_224), initialized from ImageNet-1K weights.
• Input: one 224 x 224 RGB histopathology patch, normalized with ImageNet mean and standard deviation.
• Output: seven independently parameterized sigmoid components whose sum is a bounded SSA vote-count estimate in (0, 7).
• Single-view SSA score: predicted count divided by 7.
• Four-view SSA score: mean count over rotations of 0, 90, 180, and 270 degrees, divided by 7.
• Majority-class threshold: score greater than 0.5 gives SSA; otherwise HP.
• Parameters: 22,586,759.

The study reports a five-seed result, not one uniquely selected final checkpoint. No checkpoint is designated as universally best.

Checkpoints and fixed-split results

Across the five seeds, four-view inference achieved 0.9588 +/- 0.0014 AUROC and 90.15 +/- 0.65% accuracy (mean +/- sample standard deviation) on the released 977-image fixed test partition. These aggregate values describe the five-checkpoint set and must not be attributed to an individual checkpoint.

The five-seed mean AUROC of 0.9588 is the highest documented mean among the identified studies explicitly reporting AUROC on the released slide-separated 2,175/977 partition. This is the state-of-the-art result under that fixed MHIST evaluation protocol. Comparisons are restricted to that protocol; cross-validation and frozen-encoder probing results are separate.

Inference

modeling_mhist.py is inference-only. It defines the architecture, loads a compatible locally held safetensors checkpoint, applies the evaluation preprocessing, and returns the predicted vote count, normalized SSA score, and HP/SSA label. The example below assumes an authorized checkpoint is available locally; this public repository does not provide one.

from modeling_mhist import load_model, predict

model = load_model("checkpoints/seed_42.safetensors", device="cpu")
result = predict(model, "patch.png", four_view=True, device="cpu")
print(result)

The input image must already be a 224 x 224 RGB patch. The loader does not download or contain MHIST data.

Intended use and limitations

This is an experimental research model for non-clinical histopathology research. It is not a medical device and must not be used for diagnosis, patient care, or clinical decision-making. Evaluation is limited to HP/SSA classification on the released MHIST fixed split. Performance on other laboratories, scanners, stains, organs, or populations is not established.

Confidential exclusions

This repository does not contain model weights, training code, training commands, training histories, optimizer or scheduler state, MHIST images, annotations, labels, vote-count records, data splits, per-image predictions, or internal paths.

Dataset access

MHIST is not redistributed here. Each researcher must obtain the dataset directly from the official MHIST page and agree to its Research Use Agreement. No rights to MHIST are granted by this repository.

License and attribution

The authored inference code and repository documentation are released under Apache License 2.0; see LICENSE. MHIST data and MHIST-trained weights are not included and are not licensed by this repository. The architecture uses the Apache-2.0-licensed DeiT and timm implementations; see NOTICE.

Citation

See CITATION.bib for the associated manuscript and MHIST dataset citations.