Related Experiment Video
Updated: Jun 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Benchmarking Deep Learning-Based Image Retrieval of Oral Tumor Histology
Ranny R Herdiantoputri1, Daisuke Komura2, Mieko Ochi2
1Department of Oral Pathology, Tokyo Medical and Dental University, Tokyo, JPN.
Cureus
|July 16, 2024
Summary
Self-supervised learning (SSL) methods enhance content-based image retrieval (CBIR) for oral tumors. SSL models trained on oral histology images achieve high diagnostic accuracy, even with smartphone query images.
Area of Science:
- Digital Pathology
- Oral Oncology
- Artificial Intelligence in Medicine
Background:
- Oral tumors require reliable computer-assisted diagnosis due to their rarity and diversity.
- Deep neural networks have advanced content-based image retrieval (CBIR) in digital pathology.
- A CBIR system for oral pathology was lacking due to limited databases and specialized feature extractors.
Purpose of the Study:
- To compare deep learning methods as feature extractors for oral tumor histology.
- To develop a content-based image retrieval system for oral pathology.
Main Methods:
- Construction of a large CBIR database with 30 oral tumor categories.
- Comparison of various deep learning models as feature extractors.
- Validation of model generalizability using smartphone-captured query images.
Main Results:
- Self-supervised learning (SSL) methods, specifically SimCLR and TiCo, achieved the highest average AUC (0.900 and 0.897, respectively).
- Models demonstrated strong generalizability, with high mean AUCs (0.871 for SimCLR, 0.857 for TiCo) when tested with smartphone images.
- Top 10 accuracy evaluations confirmed the clinical utility of retrieved diagnostic and differential diagnostic categories.
Conclusions:
- Training deep learning models with SSL methods on site-specific image data is crucial for effective oral tumor histology CBIR.
- This approach yields diagnostically meaningful results and high performance.
- The findings support the development of advanced CBIR systems to improve histopathology diagnosis accuracy and speed, benefiting oral tumor research.
