Five-Class Classification of Cervical Pap Smear Images: A Study of CNN-Error-Correcting SVM Models

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Machine Learning in Healthcare

Background:

• Convolutional Neural Network (CNN) models and handcrafted descriptors are complex for Papanicolaou (Pap) smear image classification.
• Existing methods primarily focus on two-class classification problems.

Purpose of the Study:

• To explore a simplified system for five-class Pap smear image classification using combined binary coding.
• To evaluate the performance of different CNN models (AlexNet, VGG19, ResNet50) integrated with Support Vector Machine (SVM) for this task.

Main Methods:

• Feature extraction from AlexNet, VGG19, and ResNet50 using transfer learning.
• Reduction of the five-class problem into multiple binary sub-problems via error-correcting coding.
• Training Support Vector Machine (SVM) classifiers and combining their outputs for final prediction.

Main Results:

• VGG19-SVM showed high accuracy (80.68%) and sensitivity (80.86%) but had long training times and false negatives.
• AlexNet-SVM demonstrated superior efficiency in running speed and prediction consistency.
• The system achieved a good diagnostic ability with an Area Under the Curve (AUC) of approximately 0.95 and high specificity (93%-100%).

Conclusions:

• The AlexNet-SVM model is a viable and efficient option for clinical application in Pap smear image classification.
• Further research should focus on hyperparameter optimization and experimental design improvements.
• The simplified binary coding approach shows promise for multi-class medical image analysis.