RunicNet: Leveraging CNNs With Attention Mechanisms for Cervical Cancer Cell Classification

Area of Science:

• Computational pathology
• Medical artificial intelligence
• Cervical cancer screening

Background:

• Papanicolaou (Pap) tests are vital for cervical cancer detection but suffer from subjective interpretation and human error.
• Limitations include false negatives and false positives, impacting screening efficiency and patient outcomes.
• Artificial intelligence (AI) offers potential for automated, accurate cell classification in Pap smear analysis.

Purpose of the Study:

• To introduce RunicNet, a novel Convolutional Neural Network (CNN) with attention mechanisms for Pap smear cell image classification.
• To enhance the accuracy and efficiency of cervical cancer screening through AI-driven analysis.
• To provide a reliable AI-based first opinion to support specialist diagnostic confidence.

Main Methods:

• Development of RunicNet, a CNN architecture incorporating High-Frequency Attention Blocks, Pixel Attention, and a Gated-Dconv Feed-Forward Network.
• Training on a large dataset of 85,080 Pap smear cell images, utilizing data augmentation and class balancing.
• Evaluation on an independent testing dataset to assess performance against baseline models.

Main Results:

• RunicNet achieved a weighted F1-score of 0.78 on the testing dataset.
• Significantly outperformed baseline models: ResNet-18 (F1-score 0.53) and fully connected CNN (F1-score 0.66).
• Demonstrated superior performance in classifying Pap smear cell images.

Conclusions:

• Attention-based CNN models, exemplified by RunicNet, show significant potential to improve cervical cancer screening accuracy.
• Integration of AI systems like RunicNet can enhance early detection rates and reduce diagnostic variability.
• AI tools can augment clinical workflows, improving the reliability of Pap smear analysis.