AutoLDP: An accurate and efficient artificial intelligence-based tool for automatic labeling of digital pathology
Background: Whole-slide image (WSI) is foundational for artificial intelligence in tumor diagnosis, treatment planning, and prognosis prediction. Efficient management of WSI labels is crucial for clinical digitalization; however, manual or semiautomatic methods limit scalability. Enhancing automatic...
Saved in:
| Main Authors: | , , , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-03-01
|
| Series: | EngMedicine |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2950489925000065 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background: Whole-slide image (WSI) is foundational for artificial intelligence in tumor diagnosis, treatment planning, and prognosis prediction. Efficient management of WSI labels is crucial for clinical digitalization; however, manual or semiautomatic methods limit scalability. Enhancing automatic pathological label recognition is critical to advancing digital pathology, improving efficiency, and drive precision oncology. Methods: We developed Auto LDP, a method for automatic labeling of digital pathology, which identifies textual information used for labeling slides. The method includes four steps: identifying text position using the CRAFT model, recognizing text content using the ParSeq model, identifying slice type using the ConvNext classifier, and combining relevant information to generate a new name. The naming format is divided into four parts: pathology ID, wax block ID, staining type, and slice type. We used the accuracy and processing time to validate our method using two validation sets. Results: The AutoLDP system was 20 times faster than manual labeling. The files per minute in the solid-state drives of CRAFT + ParSeq were the highest among all methods at 136.95 in validation set 1 and 170.95 in validation set 2. We compared the proposed model with several commonly used text detection and recognition models including ABinet, CRNN, TRBA, and Vitstr. The results show that we achieved an accuracy of 97.60 % in just 87.62 s in validation set 1 with 200 cases, which was significantly better than that of the other models. In addition, the accuracy reached 96.98 % in validation set 2 with 13,667 cases, confirming the generalization ability of the model. Conclusion: In this study, we proposed a new model, AutoLDP, automates the extraction and recognition of key information from WSI, enabling standardized naming, and significantly improving labeling efficiency. This innovation supports the digital transformation of pathology and advances precision medicine. |
|---|---|
| ISSN: | 2950-4899 |