Automated segmentation of liver hemangiomas on MRI using a modified U-Net architecture

This paper presents a method for the automatic segmentation of liver hemangioma in magnetic resonance imaging (MRI) developed by the authors, based on modifications of the U-Net convolutional architecture. An experimental comparison was conducted between the baseline U-Net, U-Net++, and Attention U-Net. It was confirmed that the Attention U-Net model, utilizing a Spatial and Channel Squeeze-and-Excitation (scSE) attention mechanism and a combined loss function (Dice + Focal Loss), achieves the best performance, demonstrating a mean Dice Similarity Coefficient (DSC) of 84.65% on a private dataset. This result is comparable to state-of-the-art benchmarks, particularly when dealing with data characterized by significant class imbalance. To validate the method in conditions approximating clinical practice, a software prototype was developed featuring a graphical user interface and the capability for integration into PACS systems via the DICOMweb protocol.

1. Kacała A. et al, Evaluation of Predictive Factors for Transarterial Bleomycin–Lipiodol Embolization Success in Treating Giant Hepatic Hemangiomas // Cancers. 2025. №17(1). P. 42. Https://doi.org/10.3390/cancers17010042.

2. A survey on deep learning in medical image analysis / G. Litjens [et al.] // Medical Image Analysis. 2017. V. 42. P. 60–88.

3. Albiin N. MRI of Focal Liver Lesions // Curr Med Imaging Rev. 2012. May, 8(2). P. 107–116. Doi: 10.2174/157340512800672216. PMID: 23049491; PMCID: PMC3462338.

4. Ronneberger O., Fischer P., Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation // Medical Image Computing and Computer-Assisted Intervention (MICCAI). 2015. P. 234–241.

5. Attention U-Net: Learning Where to Look for the Pancreas / O. Oktay [et al.] // Https://doi.org/10.48550/arXiv.1804.03999.

6. Deep Residual Learning for Image Recognition / K. He [et al.] // CVPR. 2016. P. 770–778. Doi: 10.1109/ CVPR.2016.90.

7. Coarse-to-Fine Liver Tumor Segmentation using 3D U-Net / H. Wu [et al.] // Medical Image Analysis. 2023. V. 85. P. 102734.

8. Multi-modal learning with an anisotropic 3D U-Net for the segmentation of liver tumors in DCE-MRI / A. Hänsch [et al.] // Computerized Medical Imaging and Graphics. 2022. V. 95. P. 102015.

9. Improving automatic liver tumor segmentation in late-phase MRI using multi-model training and 3D convolutional neural networks / Hänsch [et al.] // Scientific Reports. 2022. V.12. P. 12262. Https://doi.org/10.1038/s41598-022-16388-9.

10. Explainable and Robust Deep Learning for Liver Segmentation Through U-Net Network / M. C. Brunese // Diagnostics (Basel). 2025. Mar 31. V. 15(7). P. 878. Doi: 10.3390/diagnostics15070878. PMID: 40218228; PMCID: PMC11989174.