Clasificación automática de células en muestras de citología oral mediante diferentes métodos de aprendizaje profundo-images
<p dir="ltr">Oral cancer has a high mortality rate due to late diagnosis, with oral exfoliative cytology being a key non-invasive tool for early detection. However, its manual analysis is subjective and prone to errors. This study evaluates the use of deep learning to classify oral c...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| منشور في: |
2025
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | <p dir="ltr">Oral cancer has a high mortality rate due to late diagnosis, with oral exfoliative cytology being a key non-invasive tool for early detection. However, its manual analysis is subjective and prone to errors. This study evaluates the use of deep learning to classify oral cytology cells into three categories: abnormal (associated with malignancy), blood (indicative of hemorrhage), and healthy. Six pre-trained models were examined: GhostNet-100. EfficientNet-Lite4, Inception-ResNet-v2, CoaTNet-0. MaxViT-Tiny, and ViT Base Patch16, and their combination through an ensemble, using the UFSC OCPap dataset with a total of 1934 images. To optimize performance, techniques such as data augmentation (rotations, flips), Focal Loss, and the SAM optimizer were implemented to handle class imbalances. The results showed that the ensemble outperformed individual models, achieving a balanced accuracy of 84.03% and an F1 of 88% for abnormal cells (sensitivity of 87%). This approach significantly improved the state-of-the-art by surpassing the F1 for blood cells by 12% compared to previous studies. The Inception-ResNet-v2 architecture stood out individually (83.31% accuracy), while GhostNet-100 offered a balance between efficiency and precision for resource-limited environments. These results validate the potential of deep learning-based systems to optimize the diagnosis of oral cancer, reducing false negatives in early lesions and providing an objective clinical support tool.</p> |
|---|