Head localization and orientation by pruned neural networks in thermal images - a method for a robust and automatic pain detection

Many patients such as dementia patients, coma patients or children are not able to communicate their pain sensations adequately. This leads in consequence to an incorrect medication. Current body attached methods such as wearables measuring ECG are often removed by dementia patients, which results in non continuous measurments. To overcome this issue, we propose a non-contact based method to quantifying pain in thermal images automotically. A first step in our approach is the determination of the head in a thermal image by binarisation and the usage of horizontal diagrams. The subsequent determination of the heads orientation by a pruned neural network is important to locate the face regions accurately. This is followed by a local integration of pixels and the extraction of the mean temperature. Subsequently, the temperatures in the facial regions are compared over time to determine pain. In order to generate a method for the head determination and its orientation we created a database consisting of 16 persons with overall 364 images for training and testing. The actual pain assessment is carried out with 22 probands (13 male, 9 female) by applying pain in the cold pressor test. Thereby, the probands have to hold a hand into icewater and rating their pain. During the tests reference systems measuring vital parameters such as ECG, HRV, SPO2 and skin conductance for verification of the results obtained by a thermal camera. We could show that it is possible to determine the pain by means of thermal images. Furthermore, we have developed a fast method for determining the head position in thermal images. Further investigations have to focus on a better quantification of the measurements and a verification in experiments with chronic pain patients.