Atrous Convolution Neural Network for Effective Burn Region Segmentation of Color Images

Abstract Background Burn injuries are one of the most severe forms of wounds and trauma across the globe. Automated burn diagnosis methods are needed to provide timely treatment to the concerned patients. Artificial intelligence is playing a vital role in developing automated tools and techniques for medical problems. However, the use of advanced AI techniques for color images based burn region segmentation is not much explored. Method In this work, we explore the use of deep learning for the challenging problem of burn region segmentation. We prepared a pixel-wise labelled new burn images dataset for segmentation and investigated the efficacy of existing state-of-the-art color images based semantic image segmentation techniques. Lately, we proposed a new convolution neural network (CNN) that uses atrous convolution for encoding rich contextual information and utilizes pre-trained model ResNet-101 for better extraction of low-level and middle-level layer features. Results The proposed approach achieves the state-of-the-art performance on the prepared burn image dataset with 77.6% of Mathews correlation coefficient (MCC) and 93.4% of accuracy. The improvement of 11.6/5.8/6.9/1.2% is observed in precision, Dice similarity coefficient, Jaccard index and specificity, in comparison to the second best performance. Conclusion In this work, we propose a CNN based novel method for performing burn-region segmentation in color images and evaluate it using newly prepared Burn Images dataset. The experimental results illustrate its effectiveness in comparison to existing approaches. Further, the proposed pixel-level segmentation method could be useful in estimating the burn surface area and burn severity in an accurate and time efficient manner.