A human-in-the-loop strategy for defect detection in pathological slide production

The quality of histopathological slide preparation is crucial for diagnostic accuracy, yet conventional quality control methods relying on subjective evaluation by pathologists suffer from low efficiency and limited objectivity. Existing approaches based on image processing and deep learning have partially improved automation but still face challenges such as incomplete defect recognition, insufficient fine-grained detection and heavy computational demands. To address these issues, this study proposes an efficient defect detection framework that integrates image processing with a human-in-the-loop deep learning strategy. The framework first employs image processing for highly sensitive preliminary screening, followed by deep learning models for more specific secondary evaluation. In addition, an active human-in-the-loop mechanism is introduced, in which uncertain samples are annotated by experts and the model is iteratively fine-tuned to continuously enhance performance. Experimental results demonstrate that the proposed framework achieves an overall AUC of 0.985 across multiple defect types. This research enables automated, large-scale, and standardized quality control of histopathological slides, offering significant potential for remote digital pathology inspection and the digital transformation of pathology laboratories.