CNN-Based Pixel Classification for Temperature Analysis in Ultrasonic Agar Phantom Imaging

This study melds B-mode ultrasound imaging with Convolutional Neural Networks (CNNs) to pioneer non- invasive temperature monitoring in hyperthermia therapy, a critical advancement over traditional, invasive techniques. Leveraging the ALOKA ProSound 6 system, this work adapts ultrasound data to be congruent with the ResNet-50 CNN architecture, employing Ranklet transforms to elevate image quality for more nuanced analysis. The CNN, crafted with the deeplabv3plus architecture, is adept at segmenting ultrasound images into distinct temperature ranges, indicative of its potential in clinical hyperthermia applications. Despite achieving a high validation accuracy of 97.7%, the research identifies the constraints of a limited dataset and manual labeling, which may introduce biases in model generalization. These results, while robust, underscore the need for a more expansive dataset and refined labeling mechanisms to ensure the CNN's efficacy and reliability in clinical scenarios. The study signifies a notable advancement in non-invasive thermal monitoring, paving the way for improved hyperthermia treatment strategies and patient outcomes.