Dual‐ and triple‐stream RESUNET/UNET architectures for multi‐modal liver segmentation

• Published in: IET image processing Vol. 17; no. 4; pp. 1224 - 1235
• Main Authors: Elghazy, Hagar Louye, Fakhr, Mohamed Waleed
• Format: Journal Article
• Language: English
• Published: Wiley 01.03.2023
• Subjects: liver segmentation; medical image segmentation; multiple‐stream; UNET
• ISSN: 1751-9659; 1751-9667
• DOI: 10.1049/ipr2.12708

Deep learning image segmentation has become an important field of interest in recent years, especially when it comes to medical images. Segmentation of medical image modalities such as magnetic resonance imaging (MRI) and computed tomography (CT) can benefit diagnosis accuracy, speed up diagnosis process, and decrease workload. The most famously used deep learning models in the medical image segmentation are the UNET‐based models, which have been repeatedly proven to provide a high percentage of accuracy in medical image segmentation. But, most of the available datasets contain a single modality and thus are not big enough to train complex architectures. Lately, it has been shown that using multiple modalities with multiple streams architectures can provide higher accuracy more than single modality with a single stream architecture. In this paper, the benefits of dual‐stream and triple‐stream architectures are demonstrated when processing multiple modalities. This work shows that dual stream can achieve dice of 0.97 on CT images and 0.89 on MRI images, while in triple stream architectures can achieve dice of 0.97 on CT images and 0.96 on MRI images. To the best of our knowledge, these are the best results to date. segmentation of liver from small dataset that contains ct and mri images, feature merging techniques are used to achieve better segmentation across both modalities mainly by building multiple stream architectures.