Skip to Content
  • Article
  • Open Access

19 November 2026

19 Pages

Residual-Aware Feature Mapping for Anomaly Detection in Pre-Analytical Medical Specimens

,
and
1
Department of Informatics and Artificial Intelligence, Zhejiang Provincial People’s Hospital, Hangzhou 310014, China
2
Department of Orthopedics and Spine Surgery, Zhejiang Provincial People’s Hospital, Hangzhou 310014, China
*
Author to whom correspondence should be addressed.
This article belongs to the Section Medical Imaging

Abstract

This study investigates reference-based few-shot anomaly detection for pre-analytical medical specimens. Spec-ResAD forms residual features by matching each query feature to a small target-class normal reference pool, constrains the residual representation, and estimates its normal distribution with a normalizing flow. The learned feature extractor, Feature Constraintor, and Feature Distribution Estimator remain fixed for an unseen target class, while one, two, or four normal target images provide inference-time reference information. The verified hospital evaluation contains a fixed test list of 4854 images and a separate 109-image normal reference pool. Because the hospital collection has image-level hemolysis, icterus, and lipemia labels but no independent pixel-level ground truth, hospital results are reported at image level. Repeated reference selections are summarized by their mean and sample standard deviation, with the repetition unit stated for each result. The evaluation also includes MVTec data, for which official pixel masks support localization metrics. The results are interpreted descriptively under the reported protocols and do not establish statistical superiority or prospective clinical utility.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.