README.md

---
license: apache-2.0
library_name: timm
pipeline_tag: image-classification
tags:
  - medical
  - computer-vision
  - image-classification
  - brain-ct
  - hemorrhage
---

# Brain CT ICH Ensemble

뇌 CT 두개내출혈(ICH) 6클래스 분류 앙상블입니다.  
EfficientNet-B4 + ConvNeXt-Small + ResNet18 확률 평균을 사용합니다.

**연구/교육용이며 임상 진단용이 아닙니다.**

## Classes

| id | name | 한글 |
|---|---|---|
| 0 | epidural | 경막외출혈 |
| 1 | intraparenchymal | 뇌실질내출혈 |
| 2 | intraventricular | 뇌실내출혈 |
| 3 | subarachnoid | 지주막하출혈 |
| 4 | subdural | 경막하출혈 |
| 5 | any | 두개내출혈 |

## Files

- `tf_efficientnet_b4_ns_jft_in1k_fold0.pt`
- `convnext_small_fb_in22k_ft_in1k_fold0.pt`
- `ich_resnet18.pt`

체크포인트는 `model_state_dict` (또는 ResNet18의 `model`) 키를 포함한 `torch.save` dict입니다.

## Usage

```python
from pathlib import Path
import torch
import timm
from huggingface_hub import hf_hub_download

REPO = "kimsungil/brain-ich-ensemble"
NUM_CLASSES = 6

def load_ckpt(filename, model_name, device):
    path = hf_hub_download(REPO, filename)
    blob = torch.load(path, map_location=device, weights_only=False)
    sd = blob.get("model_state_dict") or blob.get("model") or blob
    kwargs = dict(pretrained=False, num_classes=NUM_CLASSES)
    if "resnet" not in model_name.lower():
        kwargs.update(drop_rate=0.2, drop_path_rate=0.1)
    model = timm.create_model(model_name, **kwargs)
    model.load_state_dict(sd, strict=False)
    return model.to(device).eval()

device = torch.device("cpu")
models = [
    load_ckpt("tf_efficientnet_b4_ns_jft_in1k_fold0.pt", "tf_efficientnet_b4.ns_jft_in1k", device),
    load_ckpt("convnext_small_fb_in22k_ft_in1k_fold0.pt", "convnext_small.fb_in22k_ft_in1k", device),
    load_ckpt("ich_resnet18.pt", "resnet18", device),
]
```

입력 이미지는 학습과 같이 380×380, brain/subdural 윈도우를 사용하세요.