Introduction
Whole-slide imaging (WSI) has revolutionized histopathology, enabling the digitalization of glass slides and computational analysis through deep learning (DL). However, the presence of artifacts during slide preparation and digitization poses significant challenges for both pathologists and AI systems, hindering accurate diagnoses. In this study, we propose a DL-based model for artifact segmentation and quality control in WSI.
Material and methods
Our training dataset consists of 100 slides with diverse tissue and stain types, digitized across seven scanners to capture real-world variability. The DL model, based on DeepLabV3+ with EfficientNet-B2 encoder, is trained to segment six common artifacts: tissue folds, pen marker, ink, air bubbles, dust, and out-offocus areas. Validation involves 500 additional cases, including preclinical tissue from rats and dogs, as well as an extended range of human organs and tissues.
Results and discussion
The model demonstrates efficacy in artifact segmentation, with a pixel-level average Dice score of 0.83 and 94% accuracy in binary classification of artifact versus non-artifact tissues. At the slide-level, an AUC of 0.95 is achieved in categorizing slides into poor or good quality based on the proportion of tissue covered by artifacts.
Conclusion
Our study highlights the potential of DL-based artifact detection in WSI to streamline the quality control process in histopathology. By prescreening slide quality, our method offers a promising solution to reduce the burden of quality control processes in clinics, ultimately ensuring more efficient and accurate diagnoses. Future work will focus on further refining the model's performance and exploring its implementation in clinical practice.