Segmentation of Liver Tumours Using Artificial Intelligence

Abstract

Liver cancer is a serious health condition affecting approximately 800,000 people annually, with around 700,000 deaths worldwide each year from this disease. One of the primary treatment methods for liver cancer is the surgical removal of the tumour. Keyhole surgery, which uses small incisions instead of a large cut, offers several benefits, including shorter hospital stays, reduced risk of complications, and cosmetic advantages such as smaller scars. However, despite its advantages, keyhole surgery is less commonly performed. Navari Surgical, a Medtech startup company founded in 2021, aims to address this challenge by developing a visual aid specifically for keyhole surgeries involving liver tumour removal, benefiting both patients and the healthcare system by improving outcomes and efficiency. This Master’s thesis investigates deep learning networks for semantic segmentation of liver tumours in CT images. This includes comparing the performance of transfer learning networks with the U-Net architecture. This analysis highlights the importance of careful dataset preparation, thoughtful model selection, and hyperparameter tuning to optimise model performance. In conclusion, U-Net demonstrated the best performance compared to the transfer learning networks, especially when prioritising the Dice score over recall. The recommendation for Navari is to further develop the U-Net architecture for future advancements in tumour segmentation. Alternatively, make small adjustments to the transfer learning networks that may lead to better performance.