Deep learning methods have been successfully applied to implement medical image segmentation. It employs convolutional neural networks (CNN) to compute distinctive image features from a defined pixel-wise objective function. However, it has become a powerful tool for computer-aided diagnosis (CAD), where multi-scale fine-grained learning of segmentation targets and global contextual interdependencies have been under constant exploration and progress, they are still the main reasons for the non-robust and unrealistic segmentation of medical images. After extensive research work, we propose in this paper five new practical and advanced methods from different perspectives, aiming to alleviate and counteract the problems in current state-of-the-art methods. They are developed using U-Net as a baseline, focusing around a multi-scale approach guided by attention flow thus obtaining fine-grained features and global contextual information. The methods in this paper make references to help neural network models improve the accuracy of segmentation results, and try to promote the development of the field of wise medical care with the goal of practical applications, and work towards the co-creation of artificial intelligence to lead medical progress.

Employment has recently emerged as a critical issue in the cultivation of talents at Chinese application-oriented universities. Traditionally, an application-oriented university aims to cultivate students’ social and practical skills. In societies in high demand for innovative talents, the orientation of the training of applied talents is tilting towards the cultivation of innovative capabilities. Taking the current employment market demand as a premise, this thesis analyses the mechanism and guarantee of employment-oriented innovative talents cultivation to provide reference for the talents training mode of Chinese universities.