Effective deep learning for oral exfoliative cytology classification.

Area of Science:

• Deep Learning
• Computational Pathology
• Medical Diagnostics

Background:

• Convolutional Neural Networks (CNNs) are increasingly used in medical image analysis.
• Sharpness Aware Minimization (SAM) is an emerging optimizer known for high performance in deep learning.
• Oral exfoliative cytology requires accurate classification for diagnosis.

Purpose of the Study:

• To evaluate the performance of SAM as an optimization algorithm for CNN-based classification in oral exfoliative cytology.
• To compare SAM with traditional Stochastic Gradient Descent (SGD) using a learning rate scheduler.
• To assess the impact of SAM on classification accuracy and overfitting.

Main Methods:

• Oral exfoliative cytology slide images were tiled and labeled by an oral pathologist.
• A VGG16 CNN model was trained using both SGD and SAM optimizers.
• Experiments were conducted with and without a learning rate scheduler over 300 epochs.
• Performance was evaluated using accuracy, precision, recall, specificity, F1 score, AUC, and effect size.

Main Results:

• All optimizers demonstrated improved performance with a learning rate scheduler.
• SAM, particularly with a learning rate scheduler, achieved the highest classification performance (AUC = 0.9328).
• SAM exhibited a notable effect size for accuracy (11.2) and AUC (11.0), and showed a tendency to suppress overfitting compared to SGD.

Conclusions:

• CNNs utilizing SAM with a learning rate scheduler achieved superior classification performance in oral exfoliative cytology.
• SAM demonstrates potential as a key tool for improving primary screening in oral cytological diagnostics.
• The findings suggest SAM can enhance the reliability and efficiency of automated diagnostic systems in computational pathology.