Automatic and accurate segmentation of liver tumors is crucial for the diagnosis and treatment of hepatocellular carcinoma or metastases. However, the task remains challenging due to imprecise boundaries and significant variations in the shape, size, and location of tumors. The present study focuses on tumor segmentation as a more critical aspect from a medical perspective, compared to liver parenchyma segmentation, which is the focus of most authors in publications. In this paper, four state-of-the-art models were trained and used to compare with UNet in terms of accuracy. Two of them (namely, based on polar coordinates and Visual Image Transformer (ViT)) were adopted for the specified task. Dice similarity measure is used for the comparison. A unified baseline environment and preprocessing parameters were used. Experiments on the public LiTS dataset demonstrate that the proposed ViT based network can accurately segment liver tumors from CT images in an end-to-end manner, and it outperforms many existing methods (tumour segmentation accuracy 56%, liver parenchyma 94% Dice). The average Dice similarity measure for the considered images was found to be 75%. The obtained results seem to be clinically relevant.
Autorzy
Informacje dodatkowe
- DOI
- Cyfrowy identyfikator dokumentu elektronicznego link otwiera się w nowej karcie 10.1007/978-3-031-42941-5_44
- Kategoria
- Aktywność konferencyjna
- Typ
- publikacja w wydawnictwie zbiorowym recenzowanym (także w materiałach konferencyjnych)
- Język
- angielski
- Rok wydania
- 2023
