Automatic detection of anatomical landmarks in brain MR scanning using multi-task deep neural networks

This work involves the development of a computer method to perform automatic graphical prescription in brain MR scanning. The approach is based on multi-task deep neural networks that perform automatic detection of the anterior commissures (AC) and posterior commissures (PC) in the sagittal images,...
Ausführliche Beschreibung

This work involves the development of a computer method to perform automatic graphical prescription in brain MR scanning. The approach is based on multi-task deep neural networks that perform automatic detection of the anterior commissures (AC) and posterior commissures (PC) in the sagittal images, and automatic determination of the symmetry line in the axial images. The proposed multi-task learning architecture solves the tasks of point and angle detection simultaneously by exploiting the commonalities and differences across these tasks. This results in improved learning efficiency and prediction accuracy for the task-specific models, when compared to training individual models separately. After deriving the AC-PC line and symmetry line on the sagittal image and axial image, respectively, the corresponding scan coverage is then determined by using an image processing approach. Based on a study using a small-sized MR brain image dataset, three benefits are observed: Firstly, our proposed approach was able to perform the task well despite limited availability of training data. This is an advantage in view of the fact that training of single task models will typically encounter difficulty in convergence using a small training set. Secondly, it achieves better performance for landmarks detection in terms of smaller position and angulation shifts between the predicted results and the ground truth. Lastly, the inference can be performed faster since the two tasks are solved simultaneously using only one model. Ausführliche Beschreibung