Multi-label classification network for multi-view dataset using dependence and attention analysis
Knee injuries are among the most common injuries people get despite of their age, gender or lifestyle. Knee injuries are often diagnosed through Magnetic Resonance Imaging (MRI) and with the increasing number of injuries yearly, radiologists are more prone to human mistakes and diagnostic errors. Since early detection of knee injuries can help in reducing the symptoms on the patients which may lead to posture and walking difficulty, an automatic diagnostic and classification tool is highly requested. There are a lot of automatic machine learning and deep neural networks solutions that can be used for such applications. Given the increase interest and demand on knee injuries studies, Stanford University released the MRNet public dataset in 2018. This dataset consists of multi-view MRI images with multi-label classification target. Since the release of the dataset in 2018, most of the designed solution including the MRNet model, deals with the classification application as binary classification resulting in using around nine deep neural networks. The focus of this research is to reduce the multi-label classification complexity by using only one feature extraction network. By using the dependence and correlation among the different labels and views, a multi-label classification network is designed that out-performs the MRNet baseline model in terms of complexity and performance.
H.T. Al Ghothani, "Multi-label classification network for multi-view dataset using dependence and attention analysis", M.S. Thesis, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2022.