DeepLearning-EndoCancer: A Data Enhancement Framework Improved Endometrial Cancer Diagnostic Accuracy

Abstract

Objective This study sought to enhance the precision of endometrial lesion categorization in ultrasound imagery via a data enhancement framework base on deep learning (DL), addressing diagnostic accuracy challenges and contributing to future research. Materials and Methods Our study gathered ultrasound image datasets from 734 patients across six hospitals. We devised a data enhancement framework including Image Features Cleaning and Soften Label, validated across multiple DL models including ResNet50, DenseNet169, DenseNet201, and ViT-B. For optimal performance, we proposed a hybrid model integrating convolutional neural network (CNN) and transformer architectures to predict lesion types. Results The implementation of our novel strategies resulted in a substantial accuracy enhancement in the model. The final model achieved an accuracy of 0.809 and a macro-AUC of 0.911, underscoring DL's potential in endometrial lesion ultrasound image classification. Conclusion We successfully developed a data enhancement framework to accurately classify endometrial lesion in ultrasound images. The integration of anomaly detection, data cleaning, and soften label strategies enhanced the model's comprehension of lesion image features, thereby boosting its classification capacity. Our research offers valuable insights for future studies and lays the foundation for the creation of more precise diagnostic tools.