Using U-Net to Semantically Segment Left Atrium in CT Images

• Main Authors: Pan, Yo-Ming, 潘友明
• Other Authors: Lu, Horng-Shing
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/zm8v9g

碩士 === 國立交通大學 === 統計學研究所 === 106 === The U-Net deep learning framework may effectively segment the left atrium, but to date this architecture has not been thoroughly experimented with regard to three important factors. These include the effect of image transformation, the effect of transfer learning, and the effect of data splitting strategy (i.e., the approach to assigning data to training, validation, and test sets). The experiments in this study confirmed that transfer learning may effectively better the model performance, and likewise image deformation may also uplift the predicted Intersection over Union (IoU) by 1~5%. Meanwhile, the pooling data splitting strategy, i.e., firstly pooling all images together and then randomly allocating them into the training/validation/test sets, may yield better results. Still, an alternative data splitting strategy that firstly randomly assigning patients into three groups and allocating their data respectively into the training/validation/test set, may be closer to the reality.