Related Experiment Video
Updated: Sep 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Automatic labels are as effective as manual labels in digital pathology images classification with deep learning
Niccolo Marini1, Stefano Marchesin2, Lluis Borras Ferris1
1Information Systems Institute, University of Applied Sciences Western Switzerland (HES-SO Valais), Sierre, Switzerland.
Journal of Pathology Informatics
|September 2, 2025
Summary
Automatic labels can train deep learning models for whole slide image classification effectively, even with up to 10% noise. This finding supports using automated labeling in digital pathology for robust AI development.
Area of Science:
- Digital pathology
- Artificial intelligence in medicine
- Biomedical data analysis
Background:
- Deep learning (DL) algorithms require large labeled datasets for biomedical analysis.
- Manual data labeling by medical experts is a significant bottleneck in training DL models.
- The reliability of automatically generated labels for DL training remains unclear.
Purpose of the Study:
- To investigate the circumstances under which automatically generated labels can be used to train DL models for whole slide image (WSI) classification.
- To determine the acceptable level of noisy labels in automated datasets before performance degradation.
- To compare the efficacy of automatic versus manual labels in WSI classification tasks.
Main Methods:
- Utilized multiple DL architectures, including convolutional neural networks and vision transformers.
- Trained models on 10,604 whole slide images across three use cases: celiac disease (binary), lung cancer (multiclass), and colon cancer (multilabel).
- Assessed model performance with varying percentages of noisy labels, including those generated by the Semantic Knowledge Extractor Tool.
Main Results:
- A performance drop-off was observed when the percentage of noisy labels exceeded 10%.
- Models trained with up to 10% noisy automatic labels achieved high F1-scores: 0.906 (celiac disease), 0.757 (lung cancer), and 0.833 (colon cancer).
- Automatic labels, when generated within the 10% noise threshold, demonstrated performance comparable to manual labels.
Conclusions:
- Automatic labeling is a viable and effective strategy for training DL models for WSI classification, provided label noise is controlled below 10%.
- The Semantic Knowledge Extractor Tool can generate reliable automatic labels suitable for DL model training in digital pathology.
- This research paves the way for more efficient and scalable AI development in analyzing biomedical images.
