Scholarlink

Domain shift is one of the main obstacles to the safe deployment of machine learning (ML) models. In cytology, whole slide images (WSIs) can have various characteristics due to the diversity of preparation protocols and scanners. To prevent domain shift, we propose a generic model that identifies the origin of a WSI, by detecting the preparation and the scanner used to produce it. Depending on the task, preparation or scanner detection, different WSI representations are suitable. We introduce a two-branch architecture to handle these representations. For scanner detection, we propose a method inspired by forensic research on device identification, and demonstrate its effectiveness with WSIs.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Ensuring the Origin of Cytological Whole Slide Images Through Preparation and Scanner Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Paul Barthe,
  • Romain Brixtel,
  • Mathieu Fontaine,
  • Arnaud Renouf,
  • Sébastien Bougleux,
  • Olivier Lézoray

摘要

Domain shift is one of the main obstacles to the safe deployment of machine learning (ML) models. In cytology, whole slide images (WSIs) can have various characteristics due to the diversity of preparation protocols and scanners. To prevent domain shift, we propose a generic model that identifies the origin of a WSI, by detecting the preparation and the scanner used to produce it. Depending on the task, preparation or scanner detection, different WSI representations are suitable. We introduce a two-branch architecture to handle these representations. For scanner detection, we propose a method inspired by forensic research on device identification, and demonstrate its effectiveness with WSIs.