Achieving enhanced diagnostic precision in endometrial lesion analysis through a data enhancement framework.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gynecological pathology

Background:

• Accurate classification of endometrial lesions in ultrasound images is crucial for patient diagnosis and treatment.
• Existing diagnostic methods face challenges in precision and consistency.
• Deep learning (DL) offers potential for improving image analysis accuracy.

Purpose of the Study:

• To enhance the precision of endometrial lesion categorization in ultrasound images using a DL-based data enhancement framework.
• To address diagnostic accuracy challenges in classifying endometrial lesions.
• To contribute to future research in medical image analysis and diagnostic tools.

Main Methods:

• Collected ultrasound image datasets from 734 patients across six hospitals.
• Developed a data enhancement framework involving image feature cleaning and soften label techniques.
• Validated the framework across multiple DL models (ResNet50, DenseNet169, DenseNet201, ViT-B) and created a hybrid CNN-Transformer model.

Main Results:

• The novel data enhancement strategies significantly improved model accuracy.
• The ensemble model achieved an accuracy of 0.809 and a macro-area under the ROC curve of 0.911.
• Demonstrated the potential of DL for accurate endometrial lesion ultrasound image classification.

Conclusions:

• Successfully developed a data enhancement framework for accurate endometrial lesion classification in ultrasound images.
• Anomaly detection, data cleaning, and soften label strategies improved the model's understanding of lesion features and classification capacity.
• The research provides insights for future studies and the development of more precise diagnostic tools.