RadSight-4B

<h2 align="center"><b>RadSight: Towards Perceptually Reliable Multimodal Radiology Image Understanding</b></h2> <p align="center"> <a href="https://arxiv.org/abs/2607.22293" target="_blank">📄 Paper</a> &nbsp;&nbsp; <a href="https://huggingface.co/unstoppableljq/RadSight-4B" target="_blank">🤖 RadSight-4B</a> &nbsp;&nbsp; <a href="https://huggingface.co/unstoppableljq/RadSight-8B" target="_blank">🤖 RadSight-8B</a> &nbsp;&nbsp; <a href="https://github.com/alibaba-damo-academy/damo-RadSight/" target="_blank">💻 GitHub</a> </p>

RadSight

Research use only. RadSight is not a medical device and must not be used for autonomous diagnosis, triage, treatment decisions, or patient-facing medical advice. Model outputs may be incomplete, incorrect, or hallucinated and must be reviewed by qualified healthcare professionals.

RadSight is a perception-driven medical multimodal large language model (MLLM) for unified understanding of 2D radiology images and native 3D CT volumes. It uses modality-specific 2D and 3D visual encoders and a shared language-model interface to support tasks ranging from fine-grained visual perception to clinical diagnosis and radiology report generation.

RadSight is trained with a four-stage progressive curriculum:

1. visual-language alignment;
2. fine-grained visual perception;
3. clinical diagnosis;
4. diagnostic interpretation.

The model is designed to learn explicit visual evidence—such as lesion attributes and spatial correspondence—before producing higher-level diagnostic or report-level outputs.

Model family

The 4B and 8B names refer to the language-backbone scale. The complete multimodal checkpoints also include visual encoders and projectors; therefore, the total parameter counts displayed by the Hugging Face interface may be larger than the variant names.

Model sources

• Code: alibaba-damo-academy/damo-RadSight
• RadSight-4B: Hugging Face checkpoint
• RadSight-8B: Hugging Face checkpoint

Quick start

1. Install the project

The released checkpoints use custom RadSight model and preprocessing code. They are not intended to be loaded as a standard text-only Transformers model.

git clone https://github.com/alibaba-damo-academy/damo-RadSight.git
cd damo-RadSight

conda create -n radsight python=3.10 -y
conda activate radsight

pip install torch==2.7.0 torchvision==0.22.0 \
  --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

CUDA 12 or later is recommended. Building Flash Attention 2 requires a CUDA toolkit compatible with the installed PyTorch version.

2. Download a checkpoint and the 2D visual encoder

# Choose one RadSight variant.
hf download unstoppableljq/RadSight-4B \
  --local-dir ./weights/RadSight-4B

# For RadSight-8B, use:
# hf download unstoppableljq/RadSight-8B \
#   --local-dir ./weights/RadSight-8B

hf download DAMO-NLP-SG/VL3-SigLIP-NaViT \
  --local-dir ./weights/VL3-SigLIP-NaViT

3. Configure the visual-encoder path

Before loading the checkpoint, edit its config.json and set vision_encoder to the local SigLIP-NaViT directory:

{
  "vision_encoder": "/absolute/path/to/weights/VL3-SigLIP-NaViT"
}

This step is required because the released configuration may contain an environment-specific path. Using a local path also prevents unexpected network access while the model is being loaded.

4. Run inference

The following example supports both 2D images and 3D CT volumes. Set modal, visual_input, and the conversation placeholder consistently.

import torch

from radsight.model import load_pretrained_model
from radsight.mm_utils import (
    get_model_name_from_path,
    load_3D,
    load_images,
)
from radsight.model.processor import RadSightProcessor


model_path = "./weights/RadSight-4B"
model_name = get_model_name_from_path(model_path)

tokenizer, model, image_processor, context_len = load_pretrained_model(
    model_path,
    None,
    model_name,
    device_map={"": "cuda:0"},
)
processor = RadSightProcessor(image_processor, tokenizer)
model.config.use_token_compression = False

# ----- Option A: 2D image -----
modal = "image"
visual_input = load_images("./example_xray.jpg")
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {
                "type": "text",
                "text": "Please generate a radiology report for this image.",
            },
        ],
    }
]

# ----- Option B: 3D CT volume -----
# modal = "volume"
# visual_input = load_3D("./example_ct.nii.gz")["image"]
# conversation = [
#     {
#         "role": "user",
#         "content": [
#             {"type": "video", "num_frames": 12},
#             {
#                 "type": "text",
#                 "text": "Please generate a radiology report for this CT scan.",
#             },
#         ],
#     }
# ]

inputs = processor(
    images=[visual_input],
    text=conversation,
    merge_size=1,
    modal=modal,
    return_tensors="pt",
)
inputs = {
    key: value.cuda() if isinstance(value, torch.Tensor) else value
    for key, value in inputs.items()
}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=False,
        modals=[modal],
        max_new_tokens=8192,
        use_cache=True,
        pad_token_id=tokenizer.eos_token_id,
    )

output = tokenizer.batch_decode(
    output_ids,
    skip_special_tokens=True,
)[0].strip()
print(output)

For reproducible evaluation, use deterministic decoding (do_sample=False) and the same prompts, preprocessing, and output parsing rules as the corresponding benchmark.

Preprocessing

For 2D images, the default training resolution is 448 × 448.

For 3D CT volumes:

• volumes are resampled to 1 × 1 × 5 mm spacing;
• Hounsfield units are clipped to [-1000, 1000] and normalized to [0, 1];
• volumes are cropped or padded to 96 × 256 × 384; and
• the same orientation, spacing, intensity, and crop/pad conventions should be preserved during evaluation.

Differences in preprocessing can substantially affect spatial grounding, anomaly detection, and report-generation results.

Training procedure

The checkpoints were trained in BF16 with a progressive curriculum. The training implementation uses PyTorch, Transformers, DeepSpeed, Flash Attention 2, gradient checkpointing, and distributed training. Refer to the project repository and paper for complete stage-specific hyperparameters.

Citation

If you use RadSight in your research, please cite:

@article{liu2026radsight,
  title   = {RadSight: Towards Perceptually Reliable Multimodal Radiology Image Understanding},
  author  = {Liu, Jianqin and Cao, Weiwei and Chang, Wanxing and Yuan, Ruifeng
             and Shi, Bowen and Zheng, Zhilin and Zhang, Xianjie and Zhang, Ling
             and Wang, Peng and Zhang, Jianpeng},
  journal = {Preprint},
  year    = {2026}
}

The paper URL and bibliographic record will be updated after the preprint is publicly available.

Acknowledgements

RadSight builds on open-source work including VideoLLaMA3, Qwen3-VL, MONAI, nnU-Net, TotalSegmentator, and RADAR. We thank the creators and maintainers of the public medical datasets used in this research.

Contact

For research questions, please contact [email protected].