Semi-automatic segmentation of brain lesions/tumours from CT head scans

Abstract

Segmentation brain lesions and tumours from medical scans is a critical medical imaging challenge that enables timely treatment planning and accurate diagnosis. However, radiologists’ manual segmentation takes a long time and is prone to error, so effective semi-automated approaches must be established. This project utilises state-of-the-art (SOTA) deep learning models to investigate a semi-automatic method for brain lesion and tumour segmentation. These models address the unique challenges provided by CT scans, such as low contrast and subtle features, while improving segmentation efficiency and accuracy through the use prompt design of pre-trained architectures. An algorithm is developed and evaluated using Python and deep learning frameworks such as PyTorch and TensorFlow to expedite the segmentation process. This method aims to reduce radiologists’ workloads, improve patient care, and speed up turnaround times by delivering accurate and effective segmentation solutions. The findings demonstrate how SOTA models can enhance clinical procedures and improve lesion and tumour segmentation accuracy.