Predicting Ethanol Drinking Status from Nonhuman Primate Liver Histology: A Proof of Concept for Parameter-efficient Fine-tuning in Few-Shot Vision-Language Modeling
Amirmohammad Mohammadi, Rupak Khadka, Kathleen A. Grant, Mary Lauren Benton, and Joshua Peeples
Few-shot classification of drinking status from rhesus macaque liver whole-slide images in the Monkey Alcohol Tissue Research Resource (MATRR). The pipeline uses frozen CONCH image and text encoders, multiple instance learning (MIL), and optional visual adapters or text residuals.
The whole-slide image dataset used in this study is not included in this repository. Data may be made available upon reasonable request, subject to the applicable data access requirements and approvals.
pip install -r requirements.txtCONCH weights are required at models/CONCH/pytorch_model.bin. A CUDA-capable GPU
is recommended for whole-slide feature extraction. environment.yml records the
Conda environment used for development.
demo.py # Feature extraction, training, and evaluation
utils/ # Models, losses, data loading, and visualization
The subject metadata in demo.py specifies the 37-animal cohort. Slide filenames
must include the cohort and subject ID, for example
Coh7a_10091_Nx_Liver_HE.svs. Preprocessing selects the left half of each slide,
filters tissue patches, and extracts 512-dimensional CONCH features. Physical
patches are 224 × 224 pixels at level 0; the encoder's supplied transform resizes
them to 448 × 448. Text inputs use CONCH's 128-position tokenization.
For a 4-shot experiment with an adapter bottleneck of 128:
python demo.py --model abmil_adapter --reduction 4 \
--shots 4 --iterations 10 --seeds 3 --epochs 250 --lr 1e-3 --patience 50Use --reduction 64 for an adapter bottleneck of 8. The MIL attention dimension
is controlled separately by --abmil_d and defaults to 64. Model choices include
Max Pool (max_pool), ABMIL (zsmil), ABMIL Adapter (abmil_adapter), and
TaskRes (taskres). Text prompts use mean pooling, as described in the paper.
Each split assigns distinct animals to training, validation, and testing, with the requested number of animals per class in both training and validation. For this cohort, 2-shot uses 4 training, 4 validation, and 29 test animals; 4-shot uses 8 training, 8 validation, and 21 test animals. The same saved splits are used for all methods. Early stopping selects the lowest validation loss.
New experiments default to outputs/experiments/; --results_root selects
another directory. Classification results are stored under raw/.
Each experiment records its settings,
source hashes, metrics, predictions, checkpoints, learning curves, and ROC curves.
Attention-based models also produce whole-slide visualizations unless
--skip_xai is specified.
Feature caches are stored in outputs/features_conch_v2/ with extraction
metadata. Saved subject splits are reused across methods. Configuration checks
prevent incompatible experiments from overwriting existing results.
Reported metrics are balanced accuracy, ROC AUC, and F1 score. Summary files contain the mean and standard deviation across splits and seeds. Attention maps show the weight assigned to each tissue patch during slide classification.
This source code is licensed under the license found in the LICENSE file in the root directory of this source tree.
This product is Copyright (c) 2026 A. Mohammadi, R. Khadka, K. A. Grant, M. L. Benton, and J. Peeples. All rights reserved.
If you use this work, please cite:
@article{mohammadi2026predicting,
title={Predicting Ethanol Drinking Status from Nonhuman Primate Liver Histology: A Proof of Concept for Parameter-efficient Fine-tuning in Few-Shot Vision-Language Modeling},
author={Mohammadi, Amirmohammad and Khadka, Rupak and Grant, Kathleen A. and Benton, Mary Lauren and Peeples, Joshua},
journal={arXiv preprint arXiv:2611.17105},
year={2026}
}