Tracking tumors via convolutional neural networks in ultrasound images

Abstract

In certain medical applications, ranging from real-time surgical monitoring to noninvasive diagnosis, it is vital to accurately identify and localize tumors in ultrasound images. Despite advances in deep learning (DL), there are still challenges in adapting these models for specific tasks in fields where data unavailability and the complexity of medical imaging are common. This study presents a self-monitoring approach for tumor tracking using ultrasound images. Additionally, the MATLAB Ultrasound Toolbox (MUST) was used to simulate various properties of the ultrasound images and to employ Optuna for hyperparameter optimization. Also, add the RUS-6000A portable ultrasound machine and the ATS-539 phantom allows the performance of the model to be evaluated in more realistic conditions. This methodology aims to improve the accuracy of tumor detection using limited data. The experimental results suggest that the model maintains relatively constant performance under different image contrast conditions, demonstrating its adaptability and reliability. However, it also suggests that additional tests should be performed to validate its generalization in real-world cases.