OralEpitheliumDB: A Dataset for Oral Epithelial Dysplasia Image Segmentation and Classification.

Area of Science:

• Biomedical Imaging
• Computational Pathology
• Oral Oncology

Background:

• Early diagnosis of potentially malignant disorders like oral epithelial dysplasia is crucial for preventing oral cancer.
• Current research is hindered by the lack of publicly accessible, annotated datasets for oral dysplasia histological images.
• This limitation impedes the development and validation of computational algorithms for auxiliary diagnostic tools.

Purpose of the Study:

• To introduce a novel, annotated public dataset of oral epithelial dysplasia tissue images.
• To facilitate the improvement of computational algorithms for the automated diagnosis of oral potentially malignant disorders.
• To provide a resource for enhancing clinical applications of computer-aided diagnosis (CAD) methods.

Main Methods:

• A public dataset of 456 oral epithelial dysplasia images from 30 mouse tongues was created.
• Images were categorized by lesion grade, with nuclear structures manually annotated and validated.
• Convolutional Neural Network (CNN) models were used for image segmentation and classification, alongside machine learning algorithms.

Main Results:

• Image segmentation using the U-Net model with ResNet-50 backbone achieved an F1-score of 0.83.
• Classification using the Random Forest method yielded the highest accuracy of 94.22%.
• The results indicate that image segmentation positively influenced classification performance.

Conclusions:

• The newly introduced public dataset is a valuable resource for advancing research in oral potentially malignant disorder diagnosis.
• The findings demonstrate the potential of CNNs and machine learning algorithms in classifying oral epithelial dysplasia.
• Further research is recommended to refine segmentation and classification stages for improved automated diagnosis.