High intensity focused ultrasound (HIFU)is a noninvasive thermal therapy used for hyperthermia and ablation treatments. Temperature monitoring is important for those procedures to induce a necessary amount of thermal dose to the target area without damaging the surrounding healthy tissues. To this end, various medical imaging techniques have been proposed. Magnetic resonance imaging provides a high accuracy temperature monitoring feature. Ultrasound is a favorable medical imaging modality for thermal monitoring due to its cost-effectiveness, accessibility and non-ionizing radiation. The speed of sound and attenuation of ultrasound waves varies with the temperature, so temperature can be measured using those ultrasound physical properties. In the previous work, we developed an ultrasound thermal monitoring method for HIFU using an external ultrasound element. The system only requires simple hardware additions such as the external ultrasound sensor and computation units, providing a temperature monitoring method at a reduced cost. However, since we use only few external ultrasound sensors, the collected time of flight information is sparse. Moreover, the thermal image reconstruction highly depends on the ultrasound element locations and its accuracy could be highly deteriorated with certain sensor locations. In this paper, we propose to reconstruct thermal images using a neural network. As this method can learn the heat evolution from a large amount of data set, it could be less sensitive to the ultrasound element location. We validated the temperature image reconstruction method on a phantom study. Promising results show the feasibility of a thermal monitoring method using an external ultrasound element and deep learning reconstruction.