Effective oral cytology deep learning classifier for annotation labels with diagnostic differences
Year: 2024
Type: preprint
Abstract: Oral cytological diagnosis is challenging and often leads to differing diagnostic outcomes. Recently, diagnostic technologies for medical images using artificial intelligence (AI) have advanced significantly. However, these advancements typically apply to cases where AI training data are uniform, leaving the effectiveness of AI models with differing correct labels unclear. This study reports on the exploration of the most effective AI model for varying labels in oral cytology images, and evaluation of classification prediction impacts when less experienced physician models are incorporated. In this study, the deep learning model ResNet50 was used, and we constructed single (individual pathologist, majority voting, and probabilistic models) and multiple models (head-fusion and logit-fusion). We also assessed the impact of less-experienced physician models on AI diagnostic performance. Among the single models, the probabilistic model was the most stable, whereas the head-fusion model among multiple models achieved the highest classification accuracy. Additionally, the head-fusion model minimized the influence of the less-experienced physician models and the probabilistic model exhibited a statistically significant decline in performance. In AI diagnosis with various correct labels, the head-fusion model proved to be the most robust and effective, suggesting that applying appropriate weightings can enhance learning and improve diagnostic accuracy.
Source: Research Square
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4404189656