Related Experiment Video For Artificial intelligence
Updated: Sep 2, 2026
Isolation of Cells with Morphological and Spatial Information from Oral Submucous Fibrosis Samples by Laser Capture Microdissection
Published on: August 11, 2023
A deep learning framework to classify oral squamous cell carcinoma arising in oral submucous fibrosis using
Chimeremeze Cosmas Offurum1, Rasika Ekanayake2, Ruwan D Jayasinghe3
1Division of Applied Oral Sciences and Community Dental Care, Faculty of Dentistry, University of Hong Kong, Hong Kong SAR, China.
Objectives:
Oral submucous fibrosis (OSF) is a chronic and progressive potentially malignant disorder with an increased risk of transformation to oral squamous cell carcinoma (OSCC). Some histopathological features of OSF may complicate adequate recognition of malignancy in OSF. This study aimed to develop a lightweight deep learning framework for classifying OSF-associated OSCC from OSF using whole-slide images (WSIs).
Methods:
This retrospective study involved hematoxylin and eosin (H&E)-stained WSIs from 156 OSF and 94 OSF-OSCC cases, treated at the National Dental Hospital, Colombo, Sri Lanka. Data were allocated into training and test datasets according to an 80:20 split. A total of 4,966,521 image patches extracted from 200 WSIs were used to train and cross-validate a hybrid-operator multiple instance learning (MIL) model for classifying OSF and OSF-OSCC. The model was then tested preliminarily using 50 WSIs (20% unseen images) and 20 external photomicrographs. Discrimination, calibration, and net benefit of the model were evaluated.
Results:
The framework achieved an AUC of 0.986 (95% CI: 0.953-1.000), an AUPRC of 0.977 (95% CI: 0.935-1.000), and a Brier score of 0.034 (95% CI: 0-0.084) on the internal test WSIs. Moreover, 19 out of the 20 external photomicrographs for preliminary external testing were correctly classified by the framework. On both datasets, the recall, specificity, precision, balanced accuracy, and negative predictive value of the MIL model in classifying OSF-OSCC and OSF were 100% (95% CI: 83.2-100%), 92% (95% CI: 80.8-97.8%), 83.3% (95% CI: 62.6-95.3%), 96% (95% CI: 82-98.9%), and 100% (95% CI: 92.3-100%) respectively. Additionally, the model had potential net benefit at all evaluated probability thresholds for classifying OSF and OSF-OSCC.
Conclusion:
A lightweight deep learning model demonstrated satisfactory performance in classifying OSF-associated OSCC and OSF at internal testing and preliminary external photomicrograph testing.
Clinical Significance:
The deep learning model may serve as a diagnosis decision-support tool for classifying OSF and OSF-associated OSCC. The model may help flag suspicious cases for pathologist review following external WSI and prospective multicenter validation.