Traumatic Brain Injury Classification Based on Slice Feature Selection and Transformer

Traumatic brain injury (TBI) usually originates from primary injury, has a high rate of death and disability, and can lead to temporary or permanent neurological deficit in severe cases. Therefore, it is very important to make an accurate imaging diagnosis as early as possible. Deep learning is widely used in medical image analysis, and the use of convolutional neural networks for initial diagnosis of traumatic brain trauma has been proven effective. Most of the traumatic brain injury recognition work in recent years is to classify a single image. We propose a new approach for diagnosing traumatic brain injuries from CT images that takes advantage of the sequence relationship between slices. Unlike natural images, medical images require doctors to synthesize information from multiple slices to make an accurate diagnosis. To address this, we combine convolutional neural networks (CNN) and Transformers to extract complex features and classify brain CT images by leveraging the correlation between slices. We also introduce an embeddable slice feature selection module that further improves accuracy. We evaluated our approach on a hospital-provided dataset and achieved superior performance compared to existing methods.