# ViX-Ray — Fine-tuned Medical Vision-Language Models

Fine-tuned weights for Vietnamese chest X-ray report generation across 3 clinical tasks and 6 model architectures.

> **Best overall performance**: Qwen2-VL-7B across all 3 tasks.

---

## Tasks

| # | Task | Description |
|---|------|-------------|
| 1 | `finding` | Generate radiology **findings** from a chest X-ray image |
| 2 | `impression` | Generate the clinical **impression** (final diagnosis) from a chest X-ray image |
| 3 | `multi` | **Multi-turn dialogue** — findings → impression via conversation history |

---

## Models

| Key | Base model | Size |
|-----|-----------|------|
| `Intern` | InternVL2.5-1B | 1B |
| `Vintern` | Vintern-1B-v3.5 | 1B |
| `Qwen2B` | Qwen2-VL-2B-Instruct | 2B |
| `Qwen7B` | Qwen2-VL-7B-Instruct ⭐ | 7B |
| `MiniCPM` | MiniCPM-V-2_6 | 8B |
| `LaVy` | LaVy-Instruct | 7B |

---

## Quick Start

### 1. Install

```bash
pip install huggingface_hub transformers torch torchvision pillow
```

For Qwen models, also install:
```bash
pip install qwen-vl-utils
```

For Intern / Vintern models, also install:
```bash
pip install decord
```

For MiniCPM, pin versions:
```bash
pip install Pillow==10.1.0 torch==2.1.2 torchvision==0.16.2 transformers==4.40.0 sentencepiece==0.1.99 decord
```

---

### 2. Download a model zip

```bash
# task  : finding | impression | multi
# model : Intern | Vintern | Qwen2B | Qwen7B | MiniCPM | LaVy

huggingface-cli download presencesw/ViX-Ray <task>/<Model>.zip \
    --repo-type model --local-dir ./
```

Example — download the best model for finding:
```bash
huggingface-cli download presencesw/ViX-Ray finding/Qwen7B.zip \
    --repo-type model --local-dir ./
```

Download all models at once:
```bash
huggingface-cli download presencesw/ViX-Ray \
    --repo-type model --local-dir ./vix_ray_models
```

---

### 3. Unzip

```bash
unzip <task>/<Model>.zip -d ./models/<task>/
# result: ./models/<task>/<Model>/
```

Or in Python:
```python
import zipfile
with zipfile.ZipFile("<task>/<Model>.zip") as zf:
    zf.extractall("./models/<task>/")
```

---

### 4. Load & infer

Set `model_path = "./models/<task>/<Model>"` then use the snippet for your model family.

#### Qwen2-VL (Qwen2B / Qwen7B)

```python
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

model_path = "./models/<task>/<Model>"

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_path, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_path)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "your_image.jpg"},
            {"type": "text",  "text": "Mô tả hình ảnh X-quang ngực này."},
        ],
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs,
                   padding=True, return_tensors="pt").to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, generated_ids)]
print(processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True)[0])
```

#### InternVL / Vintern (Intern / Vintern)

```python
import torch
import torchvision.transforms as T
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer

model_path = "./models/<task>/<Model>"

model = AutoModel.from_pretrained(
    model_path, torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True, use_flash_attn=True, trust_remote_code=True
).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, use_fast=False)

MEAN, STD = (0.485, 0.456, 0.406), (0.229, 0.224, 0.225)
transform = T.Compose([
    T.Lambda(lambda img: img.convert("RGB")),
    T.Resize((448, 448), interpolation=InterpolationMode.BICUBIC),
    T.ToTensor(),
    T.Normalize(mean=MEAN, std=STD),
])

pixel_values = transform(Image.open("your_image.jpg")).unsqueeze(0).to(torch.bfloat16).cuda()
response = model.chat(tokenizer, pixel_values, "<image>\nMô tả hình ảnh X-quang ngực này.",
                      dict(max_new_tokens=512, do_sample=True))
print(response)
```

#### MiniCPM-V

```python
import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer

model_path = "./models/<task>/<Model>"

model = AutoModel.from_pretrained(
    model_path, trust_remote_code=True,
    attn_implementation="sdpa", torch_dtype=torch.bfloat16
).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

image = Image.open("your_image.jpg").convert("RGB")
msgs = [{"role": "user", "content": [image, "Mô tả hình ảnh X-quang ngực này."]}]
print(model.chat(image=None, msgs=msgs, tokenizer=tokenizer))
```

#### LaVy

```python
import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor

model_path = "./models/<task>/<Model>"

model = AutoModelForCausalLM.from_pretrained(
    model_path, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

inputs = processor(
    images=Image.open("your_image.jpg").convert("RGB"),
    text="Mô tả hình ảnh X-quang ngực này.",
    return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512)
print(processor.batch_decode(outputs, skip_special_tokens=True)[0])
```

---

## Multi-turn (Task 3)

For the `multi` task, pass conversation history between turns:

```python
# Turn 1 — findings
response1 = ...  # run inference as above

# Turn 2 — impression (append assistant turn then ask)
messages.append({"role": "assistant", "content": [{"type": "text", "text": response1}]})
messages.append({"role": "user",      "content": [{"type": "text", "text": "Kết luận bệnh gì?"}]})
response2 = ...  # run inference again with updated messages
```

See `readme/<task>_<Model>.md` for the full per-model multi-turn example.

---

## Full Model Table

| Task | Model | Base | Zip path |
|------|-------|------|----------|
| finding | Intern | InternVL2.5-1B | `finding/Intern.zip` |
| finding | Vintern | Vintern-1B-v3.5 | `finding/Vintern.zip` |
| finding | Qwen2B | Qwen2-VL-2B | `finding/Qwen2B.zip` |
| finding | Qwen7B ⭐ | Qwen2-VL-7B | `finding/Qwen7B.zip` |
| finding | MiniCPM | MiniCPM-V-2_6 | `finding/MiniCPM.zip` |
| finding | LaVy | LaVy-Instruct | `finding/LaVy.zip` |
| impression | Intern | InternVL2.5-1B | `impression/Intern.zip` |
| impression | Vintern | Vintern-1B-v3.5 | `impression/Vintern.zip` |
| impression | Qwen2B | Qwen2-VL-2B | `impression/Qwen2B.zip` |
| impression | Qwen7B ⭐ | Qwen2-VL-7B | `impression/Qwen7B.zip` |
| impression | MiniCPM | MiniCPM-V-2_6 | `impression/MiniCPM.zip` |
| impression | LaVy | LaVy-Instruct | `impression/LaVy.zip` |
| multi | Intern | InternVL2.5-1B | `multi/Intern.zip` |
| multi | Vintern | Vintern-1B-v3.5 | `multi/Vintern.zip` |
| multi | Qwen2B | Qwen2-VL-2B | `multi/Qwen2B.zip` |
| multi | Qwen7B ⭐ | Qwen2-VL-7B | `multi/Qwen7B.zip` |
| multi | MiniCPM | MiniCPM-V-2_6 | `multi/MiniCPM.zip` |
| multi | LaVy | LaVy-Instruct | `multi/LaVy.zip` |

Per-model details (installation, full inference code) are in `readme/<task>_<Model>.md`.

---

## Citation

If you use these models or the ViX-Ray dataset in your research, please cite:

```bibtex
@article{nguyen2026vix,
  title={ViX-Ray: A Vietnamese Chest X-Ray Dataset for Vision-Language Models},
  author={Nguyen, Duy Vu Minh and Truong, Chinh Thanh and Tran, Phuc Hoang and Le, Hung Tuan and Dat, Nguyen Van-Thanh and Pham, Trung Hieu and Van Nguyen, Kiet},
  journal={arXiv preprint arXiv:2603.15513},
  year={2026}
}
```