Automated Pain Severity Detection Using Convolutional Neural Network

Pain is a sensation of physical discomfort which is caused by any kind of physical injury or illness and it is one of the most crucial factor for the patient’s recovery. Although pain assessment can be done by simple observation and self-report, objectively it is a difficult task to accomplish. In recent past, several approaches have been utilized for pain recognition by researchers, However, existing state of art techniques have several drawbacks, such as using conventional handcraft feature engineering methods, which requires domain expertise and very deep convolutional neural network which are computationally expensive in terms of training, therefore requires high computational power. In this paper, we suggest that shallowing the convolutional network can also achieve competitive performance and reduce the computational burden at the same time. In view of this, we present a convolutional neural network architecture, which utilizes only three convolutional layers. Thus, having lesser parameters and computationally efficient network. Furthermore, the proposed CNN architecture has been evaluated on UNBC McMaster shoulder pain dataset. The experiment exhibits that the result of proposed CNN based approach has achieved 93.34 % overall accuracy for multiclass pain recognition. The performance signifies, that the proposed method outperformed the existing handcraft feature-based methods and gives competitive results with other deep convolutional neural network-based methods, which are computationally expensive.