A Novel Deep Learning-Based System for Real-Time Temperature Monitoring of Bone Hyperthermia

Abstract

High-intensity focused ultrasound is a form of thermal therapy that does not involve any incisions and is used for hyperthermia and removal treatments. Monitoring the temperature is essential for such operations to provide the required quantity of thermal dosage to the target region without causing damage to the strong tissues that are located in the surrounding area. In order to accomplish this goal, a number of different medical imaging approaches have been developed. Magnetic resonance imaging allows temperature readings to be taken with a high degree of precision. Because it does not emit ionizing radiation, is easily accessible, and has a low cost, ultrasound is a medical imaging modality that is regarded favourably for use in temperature monitoring. This is of particular importance in cases involving bone tumors because of the sensitive tissues that are located nearby. Due to the fact that temperature affects both the rapidity of sound and the reduction of ultrasound waves, it is possible to estimate temperature by utilizing the physical features of ultrasound. In this article, we suggest a system that uses ultrasonography and a deep learning methodology to monitor the temperature. During HIFU therapy, the system collects data from the ultrasound channels and alternates between ablation and monitoring phases. It was created with this functionality in mind. During the monitoring phase, the ultrasonic elements in the probe are in charge of receiving ultrasound pulses that have been consecutively sent from the 256 HIFU components. We train a convolutional long short-term memory computational model using ultrasonic data to generate temperature images. Magnetic resonance thermometry readings are compared to the resulting temperature images. The mean and maximum discrepancies between each picture are calculated as a means of gauging the performance of the proposed neural network. This research suggests using a neural network to recreate thermal pictures. For this purpose, we utilize the ultrasound channel data in conjunction with a CLSTM neural network to create temperature pictures. Images of temperatures are compared with those acquired using magnetic resonance thermometry. Because this approach can acquire the progression of temperature from a vast quantity of facts, it may be less sensitive to the placement of the ultrasonic element. Phantom research allowed us to verify the accuracy of the temperature image reconstruction approach. A technique of temperature monitoring that makes use of an external ultrasonic component and deep learning reconstruction appears to be feasible, according to the encouraging findings obtained.