Hybrid Transfer Learning for Classification of Uterine Cervix Images for Cervical Cancer Screening

Area of Science:

• Medical imaging analysis
• Deep learning applications in healthcare
• Computational pathology

Background:

• Transfer learning with deep pre-trained convolutional neural networks (CNNs) is widely adopted in medical imaging.
• These networks can adapt to new domains, but irrelevant features can reduce efficacy.
• Optimizing filter usage is key to improving training efficiency.

Purpose of the Study:

• To develop a novel hybrid transfer learning technique for cervical cancer detection.
• To identify relevant filters from pre-trained networks (AlexNet, VGG-16) for improved diagnostic accuracy.
• To enhance the efficiency of CNN training for medical image analysis.

Main Methods:

• Utilized AlexNet and VGG-16 to identify relevant filters for cervical cancer detection.
• Developed a hybrid transfer learning approach, initializing a new CNN with relevant filters.
• Trained the hybrid CNN on a dataset of 2198 cervix images (1090 negative, 1108 positive).

Main Results:

• The hybrid transfer learning technique achieved a high accuracy of 91.46% in detecting cervical cancer.
• Identification and utilization of relevant filters significantly improved model performance.
• The study demonstrated the effectiveness of selective filter application in deep learning for medical diagnostics.

Conclusions:

• The proposed hybrid transfer learning method offers an efficient and accurate approach for cervical cancer detection.
• Minimizing irrelevant features through relevant filter identification enhances CNN performance in medical imaging.
• This technique holds promise for improving early cancer diagnosis and patient outcomes.