ScholarWorks@성균관대학교

초록

Liver ultrasound screenings are crucial for noninvasive liver diagnostics but are often hindered by operator dependency and varying image quality across different ultrasound devices. This study introduces a novel two-step image tone augmentation method designed to enhance the robustness and accuracy of deep learning networks (DLNs) in liver ultrasound diagnostics. Our method employs both macro and micro augmentations to simulate variations in image quality, improving model performance in detecting and classifying key liver structures. Using a dataset of liver ultrasound images, we evaluated YOLOv5 for object detection and VGG16 for classification, showing significant performance gains with the proposed augmentations. In comparison with other techniques (Histogram Equalization, CLAHE, and Gamma Correction) the proposed method consistently outperformed these alternatives across all devices, highlighting its effectiveness in managing image quality variability. Additionally, our ablation study, conducted across various architectures including ResNet50, DenseNet, and MobileNet, demonstrated that deeper models benefited most from the combined augmentations, while even lightweight models showed notable performance enhancements. These results confirm that integrating both macro and micro augmentations improves DLN adaptability to varying image conditions. Future research will explore the use of Generative Adversarial Networks (GANs) for further data diversity, as well as the application of these techniques across other medical imaging modalities and tasks. © 2025 IEEE.

키워드