ANTONI-Alpha

Vision-Language Model for Computational Pathology

Resources

• Paper: OpenReview (under review)
• Code: GitHub
• Dataset: SaltySander/HISTAI-Instruct
• Data Generation Framework: Polysome
• Base Model: MedGemma-4B-IT

Authors

Computational Pathology Group RadboudUMC

Model Information

ANTONI-Alpha is a vision-language model for computational pathology that enables natural language interactions with whole slide images.

Architecture:

• Vision encoder: Prism (1280-dim tile embeddings)
• Language model: MedGemma-4B (4-bit quantized)
• Projector: Cross-attention (256 output tokens, 8 query/KV heads)
• Training: LoRA fine-tuning (rank 16, alpha 16)

Model Statistics:

• Precision: 4-bit weights + bf16 activations

Training Procedure

Stage 1: Projector Alignment

• Frozen LLM, trainable projector
• Dataset: Clean pathology reports from HISTAI-Instruct
• 35 epochs, batch size 512, learning rate 3e-4

Stage 2: Instruction Tuning

• Full model training (LoRA on LLM + projector)
• Dataset: All English instruction pairs from HISTAI-Instruct
• 21 epochs, batch size 512, learning rate 3e-5
• Optimizer: AdamW with cosine scheduler

Installation

pip install git+https://github.com/computationalpathologygroup/ANTONI-Alpha.git

Optional: Flash Attention 2

For improved performance on compatible hardware, install Flash Attention 2:

pip install flash-attn==2.8.3 --no-build-isolation

The --no-build-isolation flag allows the build process to use your installed PyTorch. Flash Attention 2 requires CUDA-capable hardware and will be used automatically if installed.

How to Use

import torch
from antoni_alpha.models.antoni_pretrained import AntoniAlphaPreTrained

# Load model
model = AntoniAlphaPreTrained.from_pretrained(
    "SaltySander/ANTONI-Alpha",
    device_map="auto",
    torch_dtype=torch.bfloat16
)

# Load slide features (Prism embeddings)
slide_features = torch.load("slide_features.pt")  # [num_tiles, 1280]
slide_latents = slide_features.unsqueeze(0)
slide_latents = slide_latents.to(next(model.projection_layer.parameters()).device)

# Inference
conversation = [{"role": "user", "content": "Describe this tissue."}]

with torch.no_grad():
    output_ids = model.generate(
        slide_latents=slide_latents,
        conversations=[conversation],
        max_new_tokens=200,
        do_sample=False,
    )

response = model.processor.batch_decode(output_ids, skip_special_tokens=True)[0]
print(response)

Input/Output

Input:

• slide_latents: Tensor [batch_size, num_tiles, 1280] (Prism embeddings)
• conversations: List of conversations in OpenAI format

Output:

• Generated text response

Note: This model requires pre-extracted Prism embeddings. It does not process raw WSI images end-to-end.

Citation

@inproceedings{moonemans2025open,
  title={Democratizing Pathology Co-Pilots: An Open Pipeline and Dataset for Whole-Slide Vision-Language Modeling},
  author={Sander Moonemans and Sebastiaan Ram and Fr{\'e}d{\'e}rique Meeuwsen and Carlijn Lems and Jeroen van der Laak and Geert Litjens and Francesco Ciompi},
  booktitle={Submitted to Medical Imaging with Deep Learning},
  year={2025},
  url={https://openreview.net/forum?id=aGPowreqPi},
  note={under review}
}

License

This model is released under the Health AI Developer Foundations License.