ViT-LiSeg: Vision Transformer-based Liver Cancer Segmentation in High-Resolution Computed Tomography Images
Aroosa Yaqoob, Department of Computer Science, Abasyn University, Islamabad, Pakistan.
Abdul Basit, Department of Computer Science, Abasyn University, Islamabad, Pakistan.
Sibtul Hassan, Department of Computer Science, Abasyn University, Islamabad, Pakistan.
Abdul Hannan, Department of Computer Science, Abasyn University, Islamabad, Pakistan.
Corresponding Author:
Aroosa Yaqoob (aroosa.yaqoob@abasynisb.edu.pk)
Abstract:
The annual incidence of liver cancer is rising, which emphasizes how important it is to accurately identify liver tumors in order to inform treatment plans. Liver tumor delineation has traditionally been accomplished by tedious manual labeling, which is an arbitrary technique prone to error. Computer-based liver tumor segmentation techniques have gained popularity as a solution to this problem. However, because tumor sizes, shapes, and picture intensities vary widely, segmentation is still difficult. In this work, we provide a new method called Liver tumor segmentation using Vision Transformer (ViT), which seamlessly integrates spatial data in place of conventional skip-connections. ViT uses multihead self-attention to operate on low-resolution feature maps and reworks the output into a feature map for concatenation later on. For liver tumor segmentation, we use a Kaggle dataset of CT images that is accessible to the public. Our suggested vision transformer-based model shows excellent accuracy; it can diagnose liver tumors from normal liver CT images with 89% accuracy. The findings of this study may help physicians, radiologists, and practitioners diagnose liver tumors in their early stages with greater accuracy and predictability.
Keywords:
Liver Tumors; Liver Cancer; Vision Transformer; CT-Images; Segmentation; Deep Learning.