Dataset	Baseline (zero-shot)	Fine-tuned	Δ
NCT-CRC-HE-100K (9-class)	53.4%	89.6%	+36.2 pp
PatchCamelyon (2-class)	~60%	85.2%	+25.2 pp
BACH ICIAR 2018 (4-class)	~40%	68.3%	+28.3 pp
Overall	53.4%	81.0%	+27.6 pp

Model Details

Base model: google/medgemma-4b-it (4B parameter multimodal LLM)
Fine-tuning method: QLoRA (4-bit base + bfloat16 adapter merge at inference)
LoRA config: r=16, α=32, dropout=0.05, target=language layers only
Training: 1 epoch, lr=2e-5, batch_size=1 + 4 grad accumulation steps
Adapter size: ~35 MB

Training Data

Dataset	Task	Classes	Samples used
NCT-CRC-HE-100K	9-class tissue classification	Adipose, Background, Debris, Lymphocyte, Mucus, Muscle, Normal, Stroma, Tumor	5 000
PatchCamelyon	Binary metastasis detection	Normal, Tumor	5 000
BACH ICIAR 2018	4-class breast cancer grading	Normal, Benign, InSitu, Invasive	~400 (full dataset)

Usage

This repo contains only the LoRA adapter weights (~35 MB). The base model (google/medgemma-4b-it) is downloaded separately and requires a HuggingFace token with access to the gated MedGemma model.

Quickest way — Kaggle demo notebook

The notebook handles everything: GPU setup, model download, adapter loading, and launches a public Gradio UI.

Load with the HistoLab app

bash

git clone https://github.com/karaditya/medgemma-histolab
cd medgemma-histolab
pip install -e .
export HF_TOKEN=hf_...
export ADAPTER_REPO_ID=karadi97/medgemma-histolab-5k
python app.py

Load manually with PEFT

python

from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel
import torch

BASE = "google/medgemma-4b-it"
ADAPTER = "karadi97/medgemma-histolab-5k"
TOKEN = "hf_..."   # needs MedGemma access

processor = AutoProcessor.from_pretrained(BASE, token=TOKEN)
base = AutoModelForImageTextToText.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, device_map="auto", token=TOKEN
)
model = PeftModel.from_pretrained(base, ADAPTER, token=TOKEN)
model = model.merge_and_unload()   # optional: bake adapter in for faster inference

Intended Use & Limitations

Intended for: Research, education, and demonstration purposes.
Not intended for: Clinical diagnosis or patient care decisions.
The model was trained on patch-level images (224×224 px) — whole-slide inference requires tiling.
BACH accuracy (68.3%) is limited by the small dataset size (~400 images total).

License

The adapter weights are released under Apache 2.0. The base model (google/medgemma-4b-it) is subject to Google's MedGemma Terms of Use.