Classification of Pap-smear cell images using deep convolutional neural network accelerated by hand-crafted features

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning in healthcare

Background:

• Pap-smear cell classification is crucial for cervical cancer screening.
• Current methods often rely on neural networks or manual feature extraction.
• Manual classification is time-consuming, costly, and prone to interobserver variability.

Purpose of the Study:

• To enhance Pap-smear cell classification performance by combining contextual and local features.
• To investigate the contribution of different feature types to classification accuracy.
• To develop a more reliable and accurate automated analysis tool for Pap-smears.

Main Methods:

• A hybrid approach combining 29 contextual features with local features learned by a neural network.
• Investigation of feature weight distribution to understand their impact on learning.
• Extensive testing on a dataset annotated by clinical experts.

Main Results:

• The combination of contextual and local features resulted in a 3.2% increase in F1-Score.
• Numerical features were found to be an important component of the learning process.
• The proposed method demonstrated improved classification performance compared to methods using only local features.

Conclusions:

• Combining contextual and local features offers a significant improvement in Pap-smear cell classification accuracy.
• This hybrid approach can lead to more reliable and efficient cervical cancer screening.
• The findings support the value of integrating domain knowledge (contextual features) with deep learning (local features).