合成数据集

合成数据集是由阿里巴巴-南洋理工大学人才计划创建的，用于评估深度神经网络中显著性方法的可解释性。该数据集包含6400个样本，每个样本是一个细胞图像，具有易于识别的特征，并与定位地面真值掩码区分开来，从而促进了对不同XAI方法的更透明评估。数据集还包括与图像数据和标签自动生成的地面真值热图，避免了手动标记热图特征的繁琐过程。此外，还引入了mabCAM作为与我们的地面真值热图兼容的热图生成方法。该数据集旨在通过提供清晰的、无歧义的特征，展示基于热图的XAI方法并不一定产生符合人类直觉的“解释”，并推动未来研究方向的发展。

This synthetic dataset was created by the Alibaba-Nanyang Technological University Talent Program to evaluate the interpretability of saliency methods in deep neural networks. It contains 6400 samples, each consisting of a cell image with easily recognizable features paired with a localization ground-truth mask, facilitating more transparent evaluation of different XAI methods. The dataset also includes ground-truth heatmaps automatically generated from image data and corresponding labels, eliminating the tedious process of manually annotating heatmap features. Additionally, mabCAM is introduced as a heatmap generation method compatible with our ground-truth heatmaps. This dataset aims to demonstrate that heatmap-based XAI methods do not always produce "explanations" aligned with human intuition by providing clear and unambiguous features, and to advance future research directions.