Xiaohan Bai, Elektroteknik

​Title: Deep Learning for Brain Tumor Segmentation: training with foreground and background bounding box areas

Presenter:  Xiaohan Bai

Examiner:  Prof. Irene Y.H. Gu

Summary of the presentation:

Deep machine learning on medical image analysis has become a popular topic since the demand for computer-aided diagnosis has increase in past decades. For MRI scans, U-net is a frequently used DL method with good segmentation results reported. 

However, it is common that annotated ground truth data is used in supervised training. For segmentation of brain tumors from MRIs, DL methods usually require annotated tumors from medical experts, which is time consuming process.

To overcome this problem, this thesis investigates a novel approach by using foreground tumor area and background tissue area specified by 2 ellipse bounding boxes as the input of a multi-stream U-net. This is then followed by a small number of annotated tumor data (less than 20 patients) for refined training.

To further improve the performance, we also study an approach where weights are added on unbalanced classes during the training. Since the US dataset is very small, we used FG-BG trained DL network by MICAII dataset followed by refined training using a small number of patients in US dataset. Experimental results and evaluation on MICAII and US datasets are included.

Everyone is welcome!

Password 490080
Kategori Studentarbete
Plats: Online
Tid: 2021-10-25 10:00
Sluttid: 2021-10-25 11:00

Sidansvarig Publicerad: to 14 okt 2021.