Pain level perception is paramount for effective diagnosis and treatment of a patient. Pain level is identified through self-report. Because it is a subjective assessment, cognitive impairment and language barriers can hinder diagnosis. Convolutional neural networks have been used in healthcare to classify images and identify structures, proving to be effective in obtaining more accurate diagnoses. In this work, we compared the performance of nine convolutional neural network architectures for pain facial expression recognition, with three optimizers: ADAM, RMSProp and SGD. The UNBC-McMaster database was used for training and testing of the models. Two types of data were used with the networks: images with and without the face extraction. The InceptionV3 model, trained with the faces extracted and with the SGD optimizer achieved the best performance, an overall accuracy of 90.69%.
|
