数据集获取

git clone https://github.com/XRad-Ulm/FunnyNodules.git

curl -L -o repo.zip https://github.com/XRad-Ulm/FunnyNodules/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

FunnyNodules

FunnyNodules: A Customizable Medical Dataset Tailored for Evaluating Explainable AI. L. Gallée, Y. Xiong, M. Beer, M. Götz (2025)

FunnyNodules is a synthetic, vision-based dataset inspired by medical image interpretation. It is designed to support systematic, controlled, and model-agnostic evaluation of explainability methods. The dataset generation framework enables high customizability, scalability, and fine-grained manipulation of factors of variation, making it well suited for benchmarking how different models reason about image features.
The FunnyNodules dataset, employed in the experiments of this study, comprises abstract nodules described by six visual attributes:

• roundness 1-round, 5-oval
• spiculation 1-none, 5-marked
• edge sharpness 1-sharp, 5-soft
• size 1-small, 5-big
• intensity 1-dark, 5-bright
• internal structure 0-absent, 1-present
  The target class is defined based on combinations of these attributes. However, a major advantage of the synthetic FunnyNodules framework is the ability to implement different scales and rules as desired (def calculate_target in dataset_generator.py).
  <em>Figure 1:</em> Full control of image generation allows indepth analysis of attribute-target reasoning.

Key features of FunnyNodules

• 🔬 Explainability-Focused: Developed for analyzing AI reasoning methods, especially attribute-based models.
• 🎛️ Controllable: Easily adjust attributes.
• 🎯 Flexible Rules: Implement your own rules for class label.
• 🏷️ Fully Labeled: Each sample comes with complete, structured annotations for all attributes, also attribute ROIs:
  </p>

<em>Figure 2:</em> Ground-truth masks are being created during image generation and enable evaluation of attention in attribute prediction.

<p align=“center”>
<img width=“310” height=“203” alt=“Attribute_oneAttributeEffectonTarget3” src=“https://github.com/user-attachments/assets/cb60f558-f7e6-4993-bb94-30eedc0f0f3f” />
<img width=“356” height=“120” alt=“Attribute_RoundnessCSEffectonTarget3(1)” src=“https://github.com/user-attachments/assets/88d1078f-e9e5-45bd-8b6f-13e68ecbc990” />
</p>

<em>Figure 3:</em> Analysis of models’’ reasoning, see test_reasoning.py. <strong>Left

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/XRad-Ulm/FunnyNodules