Deep learning architectures for object detection with applications to MRI scans

• Main Author: Ng, Tze Yang
• Other Authors: Jagath C. Rajapakse
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/76551
• Institution: Nanyang Technological University

This report presents the results detailing the effectiveness of the Single Shot Multibox Detector (SSD) [4] in detecting brain tumor objects from T1 weighted MRI scans. The report will first detail an extensive literature review of four different architectures that utilize Convolutional Neural Networks to perform object detection in images – Region with CNN Features (RCNN) [2], Fast-RCNN [3], Faster-RCNN [7] and the SSD. For each architecture, this report will elaborate on its’ architectural design, its performance in terms of speed and accuracy on the PASCAL VOC dataset, training and test times, and its’ limitations. This report will also show how each successive architecture manages to improve on the limitation of its predecessor to improve on its performance. We test implementations of the Faster-RCNN and SSD architecture on the PASCAL VOC data set and achieve a reasonable 70.9% and 70.6% mAP respectively. We also train the two architectures on brain T1 MRI images obtained from the BraTS 2018 data set to test its performance in detecting gliomas (brain tumor). These images were segmented in the Sagittal, Coronal and Axial plane. For this we achieve a 2.57% mean Average Precision (mAP) for Faster-RCNN and 24.3% mAP for SSD.