Glioma subtype classification from histopathological images using in-domain and out-of-domain transfer learning: An
Vladimir Despotovic1, Sang-Yoon Kim1, Ann-Christin Hau2,3,4,5,6,7
1Bioinformatics Platform, Department of Medical Informatics, Luxembourg Institute of Health, Strassen, Luxembourg.
Heliyon
|April 2, 2024
Summary
This study enhances glioma classification using deep learning and transfer learning on histopathology images. A novel semi-supervised method improves accuracy and reduces pathologist workload, aiding tumor detection.
Area of Science:
- Medical Imaging
- Computational Pathology
- Artificial Intelligence in Medicine
Background:
- Accurate classification of adult-type diffuse gliomas is crucial for patient treatment.
- Traditional histopathological analysis can be time-consuming and subjective.
- Deep learning offers potential for automated and objective image analysis.
Purpose of the Study:
- To compare transfer learning strategies and deep learning architectures for glioma classification.
- To investigate the effectiveness of in-domain adaptation and semi-supervised learning.
- To develop a WSI visualization tool for highlighting tumor areas.
Main Methods:
- Evaluated out-of-domain ImageNet representations and in-domain adaptation (self-supervised, multi-task learning).
- Proposed a semi-supervised approach using fine-tuned models to label unlabeled WSI regions.
- Retrained models with ground-truth and predicted weak labels, achieving high accuracy.
Main Results:
- Achieved a balanced accuracy of 96.91% and F1-score of 97.07% with the semi-supervised method.
- Demonstrated superior performance compared to standard in-domain transfer learning.
- Developed a WSI-level visualization tool generating tumor-highlighting heatmaps.
Conclusions:
- The proposed semi-supervised learning approach significantly improves computer-aided glioma classification.
- This method reduces the need for extensive manual annotation by pathologists.
- The visualization tool aids pathologists in identifying critical tumor regions within whole slide images.
