类似病理学的可解释人工智能用于前列腺癌可解释性Gleason分级
图像与视频处理
2024-10-22 v1 人工智能
计算机视觉与模式识别
摘要
前列腺癌（全球男性最常见的癌症）的恶性程度主要依据基于Gleason评分系统的组织病理学数据进行评估。尽管人工智能（AI）在准确预测Gleason分数方面显示出前景，但这些预测往往缺乏内在可解释性，可能导致对人机交互的不信任。为此，我们引入了一个包含1,015个组织微阵列核心图像的新型数据集，由由54名国际病理学家进行注释。这些注释提供了符合国际指南的Gleason分级的详细局部分描。利用这一数据集，我们开发了一个基于U-Net架构的内在可解释AI系统，其预测结果借鉴了病理学家的术语。这一方法规避了事后可解释性方法，同时保持或超越了直接用于Gleason模式分割训练的方法的性能（Dice得分：0.713 ± 0.003（基于解释训练） vs. 0.691 ± 0.010（基于Gleason模式训练）。通过在训练中采用软标签，我们捕捉了数据中固有的不确定性，即使在高异质性情境下也能在Gleason模式分割中取得优异成绩。随着本数据集的发布，我们旨在鼓励针对主观性较强的医学任务进行分割研究，并深化对病理学家推理过程的理解。
引用
@article{arxiv.2410.15012,
title = {Pathologist-like explainable AI for interpretable Gleason grading in prostate cancer},
author = {Gesa Mittmann and Sara Laiouar-Pedari and Hendrik A. Mehrtens and Sarah Haggenmüller and Tabea-Clara Bucher and Tirtha Chanda and Nadine T. Gaisa and Mathias Wagner and Gilbert Georg Klamminger and Tilman T. Rau and Christina Neppl and Eva Maria Compérat and Andreas Gocht and Monika Hämmerle and Niels J. Rupp and Jula Westhoff and Irene Krücken and Maximillian Seidl and Christian M. Schürch and Marcus Bauer and Wiebke Solass and Yu Chun Tam and Florian Weber and Rainer Grobholz and Jaroslaw Augustyniak and Thomas Kalinski and Christian Hörner and Kirsten D. Mertz and Constanze Döring and Andreas Erbersdobler and Gabriele Deubler and Felix Bremmer and Ulrich Sommer and Michael Brodhun and Jon Griffin and Maria Sarah L. Lenon and Kiril Trpkov and Liang Cheng and Fei Chen and Angelique Levi and Guoping Cai and Tri Q. Nguyen and Ali Amin and Alessia Cimadamore and Ahmed Shabaik and Varsha Manucha and Nazeel Ahmad and Nidia Messias and Francesca Sanguedolce and Diana Taheri and Ezra Baraban and Liwei Jia and Rajal B. Shah and Farshid Siadat and Nicole Swarbrick and Kyung Park and Oudai Hassan and Siamak Sakhaie and Michelle R. Downes and Hiroshi Miyamoto and Sean R. Williamson and Tim Holland-Letz and Carolin V. Schneider and Jakob Nikolas Kather and Yuri Tolkach and Titus J. Brinker},
journal= {arXiv preprint arXiv:2410.15012},
year = {2024}
}
备注
58 pages, 15 figures (incl. supplementary)