An advanced transformer framework for liver tumor segmentation using MRI images

• Published in: Biomedical signal processing and control Vol. 107; p. 107808
• Main Authors: Sivanagaraju, P., Ramana, S. Venkata, Reddy, P.V.G.D. Prasad
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2025
• Subjects: Liver Tumor; Probabilistic filter; Res2Net; Self-attention module; Transformer; Weiner filter; Self-attention module; Transformer; Liver Tumor; Probabilistic filter; Weiner filter; Res2Net
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.107808

•To introduce a novel transformer framework for Liver Tumor Segmentation Using MRI Images.•To pre-process images using “Probabilistic fused Wiener filter (PFWF)” that removes undesirable noise and enhance the quality.•To provide a “Convolutional Res2Net model (CRes2Net)” for deep features extraction to retain accurate segmented results.•To propose an advanced Multi-level Dual Aspect Self-Attention guided Transformer (MDA-SGT) for liver tumor segmentation. Pathologists are essential in the identification of liver tumors because they use criteria like nuclear pleomorphism, invasion of tissue, mitotic activity, and cell differentiation. However, manual tumor segmentation requires a lot of work and is subject to rater variability. Thus, automated approaches become more and more necessary, requiring validation based on raters’ agreement to guarantee high-quality segmentations. Due to variations in contrast, shapes, sizes, and locations, liver lesions from MRI images are difficult to segment automatically. Automated techniques are required to increase productivity and consistency in clinical settings. This work aims to create and assess automated liver tumor segmentation algorithms from medical imaging modalities, including CT and MRI scans. A new transformer framework that incorporates liver tumor segmentation has been introduced to address those issues. In order to improve image quality, the Probabilistic Fused Wiener Filter (PFWF) is first used to pre-process images from datasets. Then, these pre-processed images are used to extract features using the Convolutional Res2Net Model (CRes2Net). These extracted features then serve as input for the Multi-level Dual Aspect Self-Attention Guided Transformer (MDA-SGT) model, facilitating accurate liver tumor segmentation. The publicly accessible LiverHCC dataset, which includes multiphasic MRI scans created especially for liver cancer research, is used for evaluation purposes. The proposed model achieves an impressive Dice Score of 97.28%, demonstrating its efficacy in accurately segmenting liver tumors from MRI images. This development could lead to better clinical outcomes for patients with liver cancer by facilitating better diagnosis and treatment planning.