Automated Segmentation of Liver from Dixon MRI Water-Only Images Using Unet, ResUnet, and Attention-Unet Models

This research aimed to develop an automated liver segmentation model for Dixon MRI water-only images using deep learning. Specifically, three popular models, namely Unet, ResUnet, and Attention-Unet, were compared in terms of their segmentation performance. The results indicate that the Attention Unet model outperformed the other two models, achieving an accuracy of 99.52%, precision of 0.96, and mean IoU of 0.94. The Unet model also performed well, with an accuracy of 99.20%, precision of 0.94, and mean IoU of 0.90. On the other hand, the ResUnet model achieved the lowest performance among the three models. These findings demonstrate the potential of deep learning models in accurately segmenting the liver from Dixon MRI water-only images. The proposed Attention-Unet model, in particular, can provide improved accuracy and efficiency for clinical applications, such as surgical planning, disease diagnosis, and treatment evaluation due to its attention mechanism focusing on the relevant features and details, ignoring irrelevant information. Overall, this research contributes to the development of automated liver segmentation models, which can significantly aid in clinical decision-making and improve patient outcomes.