Improved Traumatic Brain Injury Classification Approach Based on Deep Learning

Despite medical imaging diagnosis has made significant progress, accurate imaging diagnosis of traumatic brain injury (TBI) still remains a challenging task because of the extremely complex and diverse brain images in TBI. Deep learning has been proved to be an effective way for boosting medical image analysis performance. However, the current research in this direction is limited by the lack of a comprehensive TBI image dataset. This work contributes a new CT image dataset suitable for the detection of TBI, which includes 226 (TBI/normal: 175/51) subjects, 6780 slices in a hospital with a CT scan of the head and their ground truth classifications for TBI detection purpose given by the experienced radiologist. With this dataset, we propose a novel imaging diagnosis model of TBI based on convolutional neural network (CNN) combined with recurrent neural network (RNN) and embedded squeeze-and-excitation (SE) module. Besides, we introduce transfer learning to avoid the problems of local optimization and data insufficiency. Experimental results show that our model achieves 95.9% accuracy on the classification task of predicting whether there is damage at the slice level, which is more accurate than other commonly used classification networks. We believe that our current work can help doctors make a further clinical diagnosis.