R-Super

数据集获取

git clone https://github.com/MrGiovanni/R-Super.git

curl -L -o repo.zip https://github.com/MrGiovanni/R-Super/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

<h1 align=“center”>R-Super: Learning Segmentation from Radiology Reports</h1>
<div align=“center”>
Subscribe us: https://groups.google.com/u/2/g/bodymaps
<p align=“center”>
<img src=“documents/r_super_pdac-8.gif” width=“600”/>
</p>

Abdominal CT datasets have dozens to a couple thousand tumor masks. In contrast, hospitals and new public datasets have tens/hundreds of thousands of tumor CTs with radiology reports. Thus, we ask: how can radiology reports improve tumor segmentation?

We present R-Super, a training strategy that transforms radiology reports (text) into direct (per-voxel) supervision for tumor segmentation AI. Before training, we use LLM to extract tumor information from radiology reports. Then, R-Super introduces new loss functions (Volume Loss & Ball Loss), which use this extracted information to teach the AI to segment tumors that are coherent with reports, in terms of tumor count, diameters, and locations. R-Super can train AI on large-scale CT-Report datasets (e.g., Merlin) plus small or large CT-Mask datasets (e.g., AbdomenAtlas, PanTS). In comparison to traditional training with masks only, using R-Super to train with masks and reports improved the performance of tumor segmentation AI by up to +16% in sensitivity, F1, AUC, DSC and NSD.

<p align=“center”>
<img src=“documents/rsuper_abstract.png” width=“600”/>
</p>

[!NOTE]
Extension: Learning from 100,000 CT Scans and Reports, to Segment 7 Understudied Tumor Types
We have assembled more than 100,000 CT-Report pairs and trained R-Super to detect 7 tumor types missing from public CT-Mask datasets. Our preprint is available here.

Papers

<b>Learning Segmentation from Radiology Reports</b> <br/>
Pedro R. A. S. Bassi, Wenxuan Li, Jieneng Chen, Zheren Zhu, Tianyu Lin, Sergio Decherchi, Andrea Cavalli, [Kang Wang](https://radiology.ucsf

数据加载示例（图像类）

from PIL import Image
import glob, os

files = (glob.glob(os.path.join(path, "**", "*.png"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.jpg"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tif"), recursive=True))
print("图像文件数:", len(files))
img = Image.open(files[0]); print("尺寸/模式:", img.size, img.mode)

# torchvision Dataset 方式:
# from torchvision import datasets
# ds = datasets.ImageFolder(path)  # 要求 子目录=类别

目录组织与标注格式以源站说明和下载后实际文件为准。

完整仓库：github.com/MrGiovanni/R-Super