Deep learning-based approach for quality control scoring of digital pathological sections

Abstract Objective To explore the auxiliary role and application of deep learning-based artificial intelligence (AI) in quality control (QC) evaluation of digital pathological section. Methods A total of 2137 routine hematoxylin and eosin (HE) slides from Department of Pathology, the First Affiliated Hospital of Army Medical University, collected between January and December 2022, were scanned into digital slides. Based on slide evaluation standards, these digital slides were scored into four grades: A, B, C, and D. ResNet50, ResNet101, EfficientNet-B5, and Swin Transformer networks were then employed for classification learning. During model training, parameters trained on the ImageNet dataset were used as initial values, and QC data were utilized to perform secondary training and optimization of the models. A set of 429 routine HE slides was selected for model validation and deep learning. Results Among the four classification models, Swin Transformer achieved the best performance for all grades except for grade D, where ResNet50 performed optimally. Overall, the Swin Transformer demonstrated the highest performance with an accuracy rate of 0.83 and an Area Under the Curve (AUC) value of 0.88. The prediction speed reached 0.6 seconds per slide. Conclusion This study preliminarily validates that a deep learning-based auxiliary QC scoring system built on Swin Transformer performs well in terms of accuracy and timeliness for slide QC evaluation. This method can significantly enhance the efficiency of pathology slide QC and plays an important role in the development of an informative pathology department.