Training high-performance deep learning classifier for diagnosis in oral cytology using diverse annotations

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in healthcare

Background:

• Label uncertainty in medical images complicates accurate diagnosis.
• Variability among professionals impacts deep learning model performance in pathology.
• Standardizing annotations for oral exfoliative cytology is challenging.

Purpose of the Study:

• To develop an optimal convolutional neural network (CNN) for oral exfoliative cytology.
• To address label uncertainty by incorporating annotations from multiple oral pathologists.
• To evaluate different annotation strategies for training CNNs in medical image analysis.

Main Methods:

• Six whole-slide images were segmented into tiles using QuPath.
• Images were annotated by three oral pathologists, creating 14,535 labeled tiles.
• Six CNN models were trained using single-pathologist, ground truth, majority voting, and probabilistic labels.
• A ResNet50 baseline was used, with performance evaluated via cross-validation and repeated statistical testing.

Main Results:

• The probabilistic model achieved the highest area under the curve values (0.861, 0.955, 0.991) in three cases.
• The probabilistic model also demonstrated superior accuracy (0.988, 0.967) in two cases.
• Models trained with single-pathologist or ground truth labels showed low accuracy and high variability.

Conclusions:

• A CNN classifier trained with probabilistic labels, derived from multiple pathologists, is optimal for oral exfoliative cytology.
• This approach mitigates diagnostic uncertainty caused by inter-observer variability.
• The findings support the development of trusted AI solutions that integrate diverse professional diagnostic insights.