3D U-Net for automatic segmentation of breast tumours

• Main Author: Leow, Lucas Jian Hoong
• Other Authors: Cai Yiyu
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computer applications::Life and medical sciences
• Online Access: https://hdl.handle.net/10356/167400
• Institution: Nanyang Technological University

In recent years, a convergence of multitude of factors such as increasing demand for medical imaging studies, Coronavirus pandemic and general shift in interest has led to an exacerbating global shortage of radiologists. According to the World Health Organization (WHO), 22% of the world population will be over 60 years of age and in general, older people requires more medical imaging treatment. With this problem, countries around the world are looking into solutions to tackle this shortage and one such solution is the use of Artificial Intelligence. Computer Vision in the Medical Industry is increasingly becoming popular as technological advancement such as improvement in Graphics Processing Unit (GPU) has enabled more computing intensive Computer Vision Model to be trained. In this report, a 3-Dimensional (3D) Convolutional Neural Network was developed to perform image segmentation of tumour in the breast. 3D Computed Tomography Scans (CT-Scans) were utilized as training data. Firstly, data pre-processing techniques for the 3D CT-Scans were explored such as dividing volumes into smaller segments for memory optimization, normalization of 3D volumes and analysing depth of 3D CT-Scans. Thereafter, 3D U-Net architecture was explored with a focus on the design of a suitable loss function. A Tversky Cross Entropy loss function was explored to tackle the issue of data imbalance due to excessive background pixels. Lastly, model training and prediction pipeline was discussed with the prediction result of three models discussed. A hybrid model with optimized memory usage during model training was developed, with Dice Score of 89.84% obtained.