Related Experiment Video
Updated: Aug 21, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Oral epithelial cell segmentation from fluorescent multichannel cytology images using deep learning
Sumsum P Sunny1, Asif Iqbal Khan2, Madhavan Rangarajan3
1Department of Head and Neck Surgical Oncology, Mazumdar Shaw Medical Center, NH Health City, Bangalore, India; Integrated Head and Neck Oncology Program (DSRG-5), Mazumdar Shaw Medical Foundation, NH Health City, Bangalore, India; Manipal Academy of Higher Education, Manipal, Karnataka, India.
This study developed an automated pipeline for segmenting single epithelial cells (SECs) from oral cytology images. The system achieves high accuracy, enabling faster and less biased cancer screening.
Area of Science:
- Biomedical Imaging
- Computational Pathology
- Cancer Diagnostics
Background:
- Oral cancer has a high global incidence, necessitating advanced screening tools.
- Automated analysis of oral cytology images is challenging due to debris, blood cells, and cell clusters.
- Accurate single-cell detection is crucial for diagnosing atypical cells.
Purpose of the Study:
- To develop a semantic segmentation model for separating single epithelial cells (SECs) from multichannel fluorescent oral cytology images.
- To classify segmented cells and identify potential cancerous or precancerous conditions.
- To create an automated pipeline for improved oral cancer screening.
Main Methods:
- Utilized 2730 multi-channel fluorescent oral cytology images stained with Mackia Amurensis Agglutinin (MAA) and Sambucus Nigra Agglutinin-1 (SNA-1) markers, plus DAPI nuclear stain.
- Employed U-Net and modified U-Net architectures for multi-class semantic segmentation of SECs, cell clusters, and artefacts.
- Trained classification models, including a custom Convolutional Neural Network (CNN) called Artefact-Net, on segmented images.
Main Results:
- Modified U-Net achieved high Intersection over Union (IoU) scores for SEC segmentation (0.79).
- The Artefact-Net model demonstrated superior classification performance with an F1 score of 0.96 for segmented images.
- Artefact-Net outperformed InceptionV3 in classifying cell clusters, showing higher precision (0.91) and F1 score (0.91).
Conclusions:
- Established a robust pipeline for single epithelial cell (SEC) segmentation in oral cytology.
- The developed pipeline enables automated, bias-reduced early cancer detection.
- This automated approach has the potential to significantly improve oral cancer screening efficiency and accuracy.
