|
Establishing identity is a scientific identification activity that tests the characteristic reflections left by the objects in order to determine whether the objects appearing successively in a case are identical. To mine the key information of forensic medical images and accurately establishing identity, a multi-task learning network that can simultaneously perform clavicle segmentation, gender determination, and identity recognition for adult chest radiographs was built. Key issues such as multi-scale feature mining and integration in the case of multi-tasking, multi-loss function optimization combination, and sub-task interaction analysis were further investigated. For multi-scale feature mining and integration, a multi-resolution parallel sub-network structure similar to a high-resolution network was adopted. Different sub-tasks were connected to different sub-networks for output as needed, effectively improving the accuracy of single tasks. Auto-adaptive online settings of the loss functions balanced the learning convergence progress of different sub-networks and effectively improved the generalizability of the multi-task learning framework. At the same time, the interaction between different subtasks was analyzed, upon which the triplet loss function of integrating auxiliary information was proposed, which further improved the accuracy of identity determination. The results showed that compared to single-task learning, multitask learning achieved better results in the IOU(Intersection over Union) of clavicle segmentation and accuracy of gender and identity determination. Auto-adaptive settings of the loss functions achieved similar performance compared to iterative manual settings on weights of each sub-task. By adopting the triplet loss function of integrating auxiliary information, the accuracy of identity determination was further improved. This study performed well in the three tasks of clavicle segmentation, gender determination, and identity recognition, and had strong applicability in the research and development of multi-task learning.
|
