Intracerebral hemorrhage CT scan image segmentation with HarDNet based transformer

Abstract

AbstractAlthough previous studies conducted on the segmentation of hemorrhage images were based on the U-Net model, which comprises an encoder-decoder architecture, these models exhibit low parameter passing efficiency between the encoder and decoder, large model size, and slow speed. Therefore, to overcome these drawbacks, this study proposes TransHarDNet, an image segmentation model for the diagnosis of intracerebral hemorrhage in CT scan images of the brain. In this model, the HarDNet block is applied to the U-Net architecture, and the encoder and decoder are connected using a transformer block. As a result, the network complexity was reduced and the inference speed improved while maintaining the high performance compared to conventional models. Furthermore, the superiority of the proposed model was verified by using 82,636 CT scan images showing five different types of hemorrhages to train and test the model. Experimental results showed that the proposed model exhibited a Dice coefficient and IoU of 0.712 and 0.597, respectively, in a test set comprising 1200 images of hemorrhage, indicating better performance compared to typical segmentation models such as U-Net, U-Net++, SegNet, PSPNet, and HarDNet. Moreover, the inference time was 30.78 frames per second (FPS), which was faster than all en-coder-decoder-based models except HarDNet.