فهرست نشریات علمی دسترسی آزاد ایران

Background and Objective: Cervical cancer continues to be a major issue in public health, especially in areas with little resources. Early detection through effective colposcopy is crucial, but limitations like inter-observer variability hinder its effectiveness. Method:The current study explores the capability of deep learning models to predict the Swede score, which is a standardized method for assessing cervical lesions using ‘colposcopy’ morphology. We Propose two models for this purpose named as CCAM1 and CCAM2. We address the challenge of limited training data by employing a two-step data augmentation strategy and propose a simple CNN architecture (CCAM1). We also investigate a hybrid approach (CCAM2) combining VGG16 for feature extraction with ‘XGBoost’ for classification using limited clinical data. This emphasizes the potential of deep learning in predicting Swede scores, particularly when combined with data augmentation and hybrid methods. Results: Experimental results show that CCAM1 as our best performing model have the average accuracy (±standard deviation) of 91.00 ±3 for cervical cancer assessment.  Both models offer valuable functionalities as educational aid systems, particularly for less experienced colposcopists. Additionally, they can provide guidance on various aspects of the Swede score, offering insights beyond a simple abnormality classification. This comprehensive information can complement a doctor's analysis. Conclusion: Both proposed models in this study provide methods that can likely produce even better results in the future with the increase in the size and diversity of the available data. This highlights the importance of expanding data collection efforts to improve the accuracy and generalizability of deep learning models in cervical cancer assessment.