SkinCon

SkinCon数据集是由斯坦福大学皮肤病学和计算机科学部门合作开发的，包含3230张皮肤疾病图像，这些图像由皮肤病专家密集标注了48个临床概念。数据集的创建旨在提供一个跨多种疾病过程有用的医学数据集，以支持人工智能在医疗领域的应用，特别是提高模型的解释性和细粒度错误分析能力。SkinCon数据集不仅包括来自Fitzpatrick 17k的图像，还涵盖了Diverse Dermatology Images数据集的656张图像，增加了皮肤色调的多样性。该数据集的应用领域广泛，包括模型调试、基于概念的解释、概念瓶颈模型、错误分析和切片发现等，旨在解决医疗AI中的关键问题，如提高诊断准确性和模型透明度。

The SkinCon dataset was developed through a collaboration between the Department of Dermatology and the Department of Computer Science at Stanford University. It comprises 3,230 skin disease images densely annotated with 48 clinical concepts by dermatology experts. The dataset was constructed to offer a medical resource valuable across a broad spectrum of disease processes, supporting the deployment of artificial intelligence in healthcare, with a specific focus on enhancing model interpretability and enabling fine-grained error analysis. In addition to images sourced from the Fitzpatrick 17k dataset, SkinCon also includes 656 images from the Diverse Dermatology Images dataset, which improves the diversity of skin tones represented in the corpus. The dataset has broad applications in areas including model debugging, concept-based explanation, concept bottleneck models, error analysis, and slice discovery, aiming to address critical challenges in medical AI such as elevating diagnostic accuracy and improving model transparency.

搜集汇总

背景与挑战

背景概述

SkinCon数据集是一个由斯坦福大学开发的皮肤疾病图像数据集，包含3230张图像，每张由专家标注48个临床概念，旨在增强AI模型的解释性和错误分析能力。它通过整合Fitzpatrick 17k和Diverse Dermatology Images数据，增加了皮肤色调多样性，支持医疗AI在模型调试、概念解释等领域的应用，以提高诊断准确性和透明度。