Automatic segmentation of traumatic brain injury lesions on head CT scans using U-Net deep learning architecture

Traumatic brain injury (TBI) is one of the most serious neurological conditions requiring urgent intervention, as a concussion or blow to the head can lead to loss of neurological function and permanent damage. Therefore, the clinical course of TBI cases needs highly sensitive specialist follow-up in order to intervene early in adverse conditions and prevent progression. In the diagnosis of TBI, computed tomography (CT) scans play a critical role in identifying damage and guiding the treatment process. In this study, we propose a U-Net deep learning architecture combining the encoder part with VGG-19 for segmentation of TBI-related lesions on CT scans collected from Bilecik Training and Research Hospital. To improve the performance of the model, data augmentation techniques are applied to the dataset and the loss function is adjusted to prevent the model from overfitting. To improve the segmentation performance, a combination of Binary Cross Entropy (BCE) and Dice Loss is chosen as the loss function. The segmentation performance of the model is evaluated using the key metric of Dice similarity coefficient (DSC), and a segmentation score of $\mathbf{0. 8 4}$ is obtained. Experimental studies have shown that the proposed method has a remarkable performance in TBI lesion segmentation.